New Mac Studio with M5 Max and M5 Ultra | Hacker News Reader

New Mac Studio with M5 Max and M5 Ultra

apple.com

823 pointsinterpol_p561 comments

Due to pricing insanity (not that Apple prices weren’t insane before the ram/ssd shortages) I’m not in the market but I do wonder if my next computer should be a Mac Studio instead of a MBP that lives its life docked. Might be better to just run a Studio and Neo for the very few times I actually need remote capabilities.

I had the same thought, I grabbed a studio two years ago for this reason and it’s been great. 99% of the time lack of portability isn’t a concern. Every now and then (e.g. travel) I notice the limitation, but it’s not much of an inconvenience to just not do some work for a bit.

Plus remote work is getting easier and easier. There are so few instances when I'm not able to get online. If we lived in a world where hardware were getting cheaper, it might make sense to splurge. In this environment I think the Neo is perfect.

Build quality of the Neo is extremely good, I love the keyboard — it’s more tactile and reminds me of early 2010s MacBooks.

I’ll be selling my M4 MBA soon, I genuinely use the Neo more. Huge difference in typing experience.

Great repairability is a plus. It was super easy, and actually fun to open. Felt like unboxing an Apple product. Applied the thermal paste mod for $10 which works excellently; I’ve had it shortly after launch.

And I love the notchless display, even if I wished the color gamut was a bit better.

The Neo keyboard just feels nicer than even an M5 Air.

That’s what I have been doing for years, it remains in the house secured while I ssh into it from an old thinkpad. You can get air to pair it with it if you really wanna have that seamless flow, otherwise, ssh works well.

I was in the same situation, I used maxed out 15'' M3 Max MacBook Pro docked to Studio Display closed on vertical stand behind the screen. It was fine for office work, but running local LLMs would definitely overheat it. The battery started degrading purely due to heat issues. And it was audible as well.

I decided to get Mac Studio M4 Max, also all maxed out config and the cooling is so much better that I can run local LLMs like Gemma 3/4, gpt-oss 120b all day long without any heat issues or any audible fan noise. So for my use case it was the right decision. I subsequently added 15'' M5 Max MacBook Pro all maxed out to my collection and even though it is slightly faster on LLM inference (I get 100 tokens/s with Gemma 4 27b model), you just can't run LLMs longer than a few minutes. It starts overheating and gets really loud.

Weird, I’ve run LLM batch sessions for hours on my Max M3 MBP. It doesn’t get very loud, though I’m not getting anything close to 100 tok/s on a 27b model, I use a 35b MoE model just to get 90 tok/s. The fan comes on but thermally it never overheats. I do have it in a vertical closed position, though.

I've been holding out, because I think my next purchase will be a Studio with an Ultra Chip in it. I'm wanting it to be a "forever" server, so I'm holding out while I can.

Buy in 2 years, or buy now and have it last 2 years less than forever.

Then you'll always be waiting, there's always something new the industry tries to tempt you with.

I just wait for a cycle or two where the leaps and bounds are more like hops and steps. So if the M7 Ultra improves inference by 2x over the M5, but the M9 Ultra only improves by 1.2x over the M7, that's my signal to buy. Unfortunately they haven't slowed down yet.

Supply rumours are next year we see an M7 AI-focused chip with large inference performance upgrades. It's unlikely we'll see heavy upgrades in other areas. If you care about AI, it's worth waiting. If you don't, pull the trigger now. RAM constraints are likely to get worse next year. Or wait 2-3 years and prices should be back to Earth (plus newer and even better chips).

I'm waiting this out.

Yeah, but how many years until 128GB+ is attainable by mere mortals again? My 2021 home server build was 64GB of RAM. My 2025 build was 32GB and zram. :/

My desktop with 64gb of RAM has been my best performing financial asset. The value keeps rising.

If running LLMs locally matters, it’s hard to imagine a “forever machine” existing in anything less than 5-10 years, probably more. This stuff is just evolving so rapidly. Buying a “forever machine” today might be like buying a “forever GPU” in 2003.

How about business where you send in all your old devices and get back a SSD using using their memories

Laptops can't do agentic engineering. They get hot as hell and battery drains instantly. I think this will promote a switch to desktops for the next couple of years, until we have new mobile chips.

But laptops can remote into boxes that can run agents.

So a combination of a powerful desktop and a "cheap" laptop might indeed be attractive.

Are you referring specifically to agentic engineering with locally hosted models?

I dusted of my lightest computer with an M1 chip and use Tailscale to make my network virtual from anywhere. Been running a. Pi5 as a main house hub and an M1 Pro as an always on Mac. It would be nice to go all out and make a Studio a hub I can just screen share into for major compute.

Do it! I went with the Mini/Neo combo. I don't need MBP power when out and about. When at home, the Mini is all I use.

If all you want to do is remote into your desktop, Neo seems like overkill. Why not just get a $200 Chromebook and save yourself $500?

Because the screen, keyboard, and especially trackpad on a $200 Chromebook sucks?

Also: the enclosure, and the hinges.

Also it'll be truly too slow. You need to at least be able to access a shared document while a video call is going.

The funniest recurring thing when working at Google was new hires getting baited into taking a Chromebook, then not being able to switch to a Mac for like 2 years. Our team made sure new people didn't fall for that.

Because who hates themselves that much? It's the thing I touch and interact with. That's exactly the part that needs to be sturdy, smooth, and pretty. It's the facade to the beast at the other end.

This comment got me curious, so I searched for cheap laptops with mobile internet. There's at least one LTE Chromebook now for about 300 euros. This might be my next laptop if it can run Tailscale and remote desktop.

Also remoting into a Mac sucks on anything that isn't a Mac. It runs like shit when using any VNC client that isn't Apple's.

I've been thinking about this a fair bit recently.

We make a lot of price/performance compromises for having an attached screen and keyboard on our computer. That was what got me started.

Then I remembered the days of having to go to a special corner of the house to use a computer, vs now when I have a computer with me all the time. In my bag, on the sofa, on the train. Hell, I'm writing this on the work MBP while waiting for an appointment.

And you know what, I think I got more done when I went and sat in a corner of the house all those years ago. I set up an area for "computer work", and it worked really well.

I have a home office, but it's a jumble of cables going into docking stations and all sorts of weird stuff. I think if I streamline it and turn it into a proper "computer room", I might get some of that mojo back. I might even convince my partner that surrendering the home office and having a corner of the den might be good - she can watch TV while I tinker. And I won't be balancing a laptop on my knee and trying to do two things at once.

And the price/performance thing comes back in. Hmm.

There is a definite mental aspect for most WFH folks to having a space that is dedicated to work. I'm not unique in saying this, but the way I put it is "If you work from anywhere in your house, then you're always at work."

And that, from mental load standpoint, is not healthy for most folks.

Or, if you're like me then "if you work from anywhere in your house, then you're never at work."

On some days I wor from home but need to be present and the sun is out. So I take my laptop outside and work from there. I cannot do this with a desktop. Wouldn‘t the workaholic just stay inside and miss out?

Sometimes I'll just pop into a meeting from my phone and walk around. That said , I do have a laptop but only because I go to the office every so often . Otherwise, I work exclusively from my home office because that's where my home desktop computer is and I control everything on my work computer from there.

Camera on or off?

I love taking meetings while walking but since Covid everyone has their camera on making sedentary meetings mandatory.

There's a lot of value in camera-on, but for some meetings, I love being camera off, nowhere near my computer, somewhere else in the house, watering plants or sweeping the floor. Sometimes it actually has me MORE engaged with the verbal aspect of the meeting, listening better, speaking better, because there's less distraction from Slack and Hacker News...

I'm much more attentive in some meetings when my hands aren't on the keyboard. For particularly important meetings I'll get out my knitting so that I can be close by while having something that's not failing unit tests occupying my hands

Off now, but historically even at camera on companies I'd do whatever and just hold my phone when convenient for me or just shove the meeting in my pocket if I need my hands.

Obviously depends on your team culture, but try it :)

Wireless headphones can get you halfway there if they have enough range for ya to walk around the house a bit. Airpods are quite good for this.

What if there was a way to use your home computer from a very cheap laptop?

I've done that, but I chose a Framework Desktop instead. The latest Fedora is closer to Snow Leopard than anything Apple has to offer. Downside is that I still have my MBP because of the lock-in and occasionally pick it up to do computer stuff in weird places.

I got used to the hub-and-spoke model at home (previously thin terminal, server-client, etc.). Big ole desktop/server and smaller devices that (ab)use it remotely. Roam around with a smaller computer/tablet/phone. Tailscale to bind it all together.

If your computing needs line up, it's a very serviceable approach.

Yes, this is my approach now: a home server with devices connected via Tailscale. I'm starting to do more from my phone and less on the laptop.

I haven't added my iPad to the Tailnet yet but i reckon that it could become a very comfortable and productive device for me.

While I do love Fedora, I Wouldn't call it a "Snow Leopard". Around August 5th, there's been numerous regressions when it comes to AMD graphics. First a kernel regression that resulting in artifacts on the desktop. Then a linux-firmware regression with decoding AV1 videos that resulting in 4 prominent columns and constant major color shifts.

Its update policy can result in this sort of regressions in a middle of a release. One thing that's nicer thing about macOS is that a released version doesn't regress as much (even if they are buggier at release; but in that case you can just wait until a later point release to upgrade).

BRING BACK THE COMPUTER ROOM

A fathers gift

We're turning your bedroom back into the computer room

Got my hiking boots ready.

The wired network connection is just so good.

But the couch is just so comfy.

To me the idea of working on (non-docked) laptop always seemed like an idea that you would do only if there is no other possible way.

Screen is small and is only one. Ergonomics is entirely messed up. Either your screen is too low, or your keyboard is too high. Keyboards are non-ergonomic and have to be made with compromises due to height limits. Touchpad instead of mouse/trackball is compromise for many - and also stuck at one position.

And yet they have somehow spread despite number of people going on business trips not really increasing.

You can dock your laptop at work, then dock it at home, take it with you to a meeting room at work to browse supplemental materials and/or project your screen n the meeting, and use it for the occasional trip (like using it on the train to work).

It's not like you have to be on a plane for the portability to be useful.

I think the point of OP is that while laptops are fine for business trips, tethered usage is superior. My setup is a laptop. Both at home and as work I have a solid desk with an ergonomic chair. On the desk are thunderbolt docks, the docks have hooked the keyboard I like, 2 monitors on mobile stands. And that's it. I have most the advantage of a desktop while using the same machine. One cable to unplug. Super nifty.

I see it now, I misread.

Still, being able to take you laptop to a meeting to either project, or to look up related material to catch up and clarify before making a point in a meeting etc. is invaluable (no, working at a laptop during a meeting doesn't mean you're not paying attention, it often means you're cross-referencing discussions with other data live).

It sounds like you don't code much in bed or with your feet up.

A laptop is a distinct tool from a desktop and if you try to use it like a desktop, I agree it is a terrible substitute. Personally I find laptops vastly more ergonomic than a desktop, but I never program with my feet on the ground.

> code much in bed

I've found my smartphone's the best device for that. I wrote an entire programming language using my phone.

Been trying to make a handheld cyberdeck to replace the phone with something better, but it's still a long way from being real.

Can you share more about your setup? I sort of hate my laptop. I carry a ridiculously overpowered phone around with me all the time, though, so the dream is just to dock it at home or at work and just use that. It’s a shame that Apple does not make this a better experience with iPhones; I would happily shell out for the most expensive iPhone if it were a real desktop replacement. Being able to develop (for real, with a terminal) is a must though.

Android phone with termux and all the usual Linux tools. I use vim to write code, and I type with the touch screen. Not exactly peak ergonomics, but it works. For me, ability to code while lying down comfortably beats having a physical keyboard and bigger screen.

Set up wireguard in my router and now I've got remote access into my LAN. My laptop has essentially become a Linux server. I just ssh in from the phone and write code. Can also easily launch and use codex.

I have been using docking stations since 2006, and apparently if I want to buy a serious desktop that doesn't look like car tuning in a box, I have to build it myself, given the available configurations on ready to go desktops.

I went some time without a desktop computer at home. The idea was, I have a perfectly good laptop, I'll just connect it with a dock when I want to use a desktop-like device.

Never ended up happening. I didn't get much done at home at all and what I did get done was mostly on the couch; not a great environment for serious work. (I have an office where I do for-profit work so it's not an issue, but still, I like my side projects too)

Eventually I had enough decommissioned computer parts that I could assemble them and re-commission them into a working desktop. So I now have a desktop again. And I actually end up sitting at the desk working on stuff on the desktop in a way I rarely did with the laptop.

every time i try to 'streamline the computer setup' i get reminded that losing 'the desk' was the worst part / saddest moment

Is there an option to "use the remote computer just as it was local" even between two Macs? Sure, terminal/ssh is enough for most use-cases, but this is something I expected from Apple as builtin.

I can tailscale into my home network, but what next? VNC/TeamViewer/Remote Desktop setups are all mediocre imho.

Apple Remote Desktop is very usable. I have a headless mini and have managed to fool myself about which machine I’m on, over WiFi.

What are you using to connect to the headless mini to use as a primary?

I’m on a MacBook Pro. Using a combo of Terminal, Apple Remote Desktop or Screen Sharing (from desktop, command K, VNC://<IP> then connect.

It’s pretty good.

To be clear, it’s not ever supposed to be a primary. It just becomes one from time to time when I make it full screen and am not paying attention.

Astropad Workbench

I once tinkered with the game streaming set up available via Steam, so I could do it in a way where it streamed the normal complete desktop setup. I used it to operate my desktop from the laptop.

It worked relatively well, but wasn't perfect. It was as close to what you're describing as I've seen, though. This was a few years ago; something like this might work even better now.

Unfortunately there is nothing that works as nicely as Windows Remote Desktop. Probably the best option is various commercial offerings as Apple Remote Desktop is rather buggy and unstable.

Apple screen sharing works for me just fine. Matter of fact I have been driving a mac mini with it from my laptop most of the time now and just ordered a mac studio to replace the mac mini. I think powerful headless desktops are the future.

I'm writing this right now connected to my Mac Mini from my "thin client" M1 MacBook Air. Screen sharing is pretty seamless nowadays, particularly if you're on a local LAN (although it works pretty well over Tailscale as well if the WiFi is strong enough).

If you haven't tried Screen Sharing since they deprecated VNC and switched to their proprietary H.264-based protocol, its worth trying. Even YouTube videos play just fine with no noticeable lag.

Thanks! I didn't even know this existed!! :o

Well, there used to be nxhosting back in the NeXT days, which I _really_ miss.

Unfortunately, it relied on Display PostScript and Quartz née Display PDF isn't architected to allow that sort of remote display/access on a per application level.

I find Jump Desktop to be excellent. One time payment. Can connect from iPhone / iPad, too. A noticeable step up from Screen Sharing (which can’t be used from mobile device anyway). Can’t recall it ever failing to connect or flaking out as so many RDP setups seem to.

Apple Screen Sharing is probably the best when it's purely within Apple products. But if the mix is more heterogeneous, I've found Sunshine (server) and Moonlight (client) is great.

This open source software was built for lag-free gaming over a network, but I've found they're great for desktop usage too. I frequently use it to remote into my Windows tower. Feels about as responsive as if I was sitting at the desk using the tower directly.

I switched back to a desktop tower a few years ago as well. Notably, I don't use external displays with the work-issued laptop, but rather set up the laptop on a stand alongside the dual 1440p displays and have synergy for mouse and keyboard sharing.

Man I did exactly this with the PC I built last April, it was the best decision ever. I love having a dedicated space with a huge tower running everything. Two big monitors mounted to the back of the desk on arms, huge mechanical keyboard in the middle with a big ergonomic mouse pad and hand shoe mouse.

It’s a whole “thing” when I go to the computer now. And frankly, it’s made it way more fun to use. It’s kind of like enjoying the process of listening to vinyl rather than pulling up a song on Spotify

As long as we keep going back-and-forth and buying stuff, trying to figure it out, the economy will be in good shape.

You don’t need to spend all that money to get your mojo back, make the changes that you would make I.e. leave the MacBook on your desk and see if life improves.

That’s nearly what I do but on a smaller scale. My iPad Pro serves as my laptop 90% of the time, and the 10% of the time I need to actually code and test in a chromium browser I remote into a mini.

Having replaced my MacBook with a Mac Mini, I would reconsider. The MacBook is just such a _complete_ package. Great speakers, great keyboard, fantastic screen, the fingerprint sensor thingy.. Takes a lot of gear to match that

If your computer never leaves your desk, the iMac is pretty competitive with those features. Even comes with a TouchID keyboard.

I’ve had iMacs for almost 20 years. My last one was indeed my last. Without target display mode (use the Mac as a monitor), I’m ditching a perfectly good monitor. I was going to buy a Mac Studio and a good monitor to replace the iMac until the spouse reminded me that we are now retired and will spend time in a camper. So a MBP for me (and BenQ’s Mac-specific monitor), but others might do well to consider a Mac Mini/Studio.

iMacs are great for a lot of use cases, but my image of the typical HN user would prefer to keep the monitor separate.

It's like, you pay laptop prices for laptop performance but without laptop portability or desktop modularity. This only sorta made sense back when Apple had no "prosumer" desktop offering unless you wanted a cheap used Mac Pro (which is what I ended up doing at some point).

Not an option if you're trying to drive 3 decent screens.

Macbook (base) yes, but my MBP (M3 Max) is driving 4 screens for me without issue.

My M1 Pro drives 4 screens no problem. The trick is that you can only attach two through one Thunderbolt port, so you have to attach the 3rd one via the other port instead of your dock. (The 4th screen is the laptop's built in screen)

Keyword is pro. Non-pro chips drive fewer screens unless you use DisplayLink. Which I did at some point, it's a hack but ok for lighter usage.

Yes but if your debate is desktop or laptop, you're probably not looking at anything but a pro.

Uh idk, I have a desktop and laptop that are both base. The laptop is an M4 Air which is overall faster than M1 Pro, especially single-core where it's even faster than M3 Max.

I've been away from Macs (2015 MBP and an M1 Pro) for a while now, but they used to have a limitation of one screen per port. Apparently driver related as according to my faulty memory a Linux install didn't have the same issue.

Are you using Apple displays or did they fix it?

They must have fixed it because I drive my two main monitors from a dock connected to one port and then my 3rd monitor from the other port. Been set up this way for at least two years now.

I've been away for about 3 years now. My thunderbolt dock couldn't drive two monitors from my Macs, and I couldn't daisy chain display ports either. A Linux ThinkPad had no problem though. Glad they've finally fixed (implemented) it.

Note that modern Macbook Pros can drive crazy screens.

A Macbook Pro M4 or M5 drives TWO of these at once, well, for equivalent of quad 4K configured as dual super ultrawides:

https://www.samsung.com/us/monitors/gaming/57-inch-odyssey-n...

Per Apple support page, MBP with M5 Max handles:

Two external displays

Two displays up to a native resolution of 8K (7680 x 4320) at 60Hz, 5K (5120 x 2880) at 120Hz, or 4K (3840 x 2160) at 240Hz over Thunderbolt or HDMI

Three external displays

Two displays up to a native resolution of 6K (6144 x 3456) at 60Hz or 4K (3840 x 2160) at 144Hz; plus one display up to a native resolution of 8K (7680 x 4320) at 60Hz, 5K (5120 x 2880) at 120Hz or 4K (3840 x 2160) at 240Hz over Thunderbolt or HDMI

Four external displays

Four displays up to a native resolution of 6K (6144 x 3456) at 60Hz or 4K (3840 x 2160) at 144Hz over Thunderbolt or HDMI

https://support.apple.com/en-us/101571

PS. Use TB4 or TB5 dock from Cal-Digit (match the TB version of your mac), or this or one of the rebranded versions of the same OEM: https://www.owc.com/solutions/thunderbolt-dock, or an Ivanky Fusion dock if you want to go nuts with up to SIX displays, Quad 6K@60Hz and Dual 4K@60Hz:

https://www.amazon.com/Thunderbolt-Monitor-Docking-Station-A...

I do remember the M1 chips being uniquely capable in this aspect compared to later chips but when I say decent displays I meant 3x 120hz 4k screens with HDR on at least one.

Last year when I was browsing the only recentish Apple silicon capable of driving that was the M3 Ultra.

Honestly the screens become more of a liability than an asset to me past 2 (including the laptop screen). I've tried. Even if I'm knee deep in work and have like 4 servers I'm monitoring, my eyes are only going to look at one screen at a time, so it's just extra window management and eye/head/mouse movement to have more screens.

I thought the question was low-end MacBook + mini vs high-end MacBook. The low-end one has all those nice things. I wouldn't sacrifice that. If it has to be MacBook + super old used mini as home server, so be it.

Except for LLM inference speed, macs suck as a server and are pretty expensive, why wouldn't you get a Linux box?

Because it's not exactly a server, more like a home PC / dev machine that I also remote into. Usually SSHing but sometimes VNC which is annoying in Linux. And the Mac mini is pretty fast. Yes a Linux box would win in terms of multicore CPU(?) or GPU if you need to run big batch processes, or a high-uptime service. My work is running those on cloud, not in my house.

Buying a Mac specifically for serving is misguided in nearly all use cases, but as an all-in-one solution they make sense. Personally I feel like the number of boxes I need to manage has an inverse correlation to my happiness.

Macs are totally “fine” for light server duty… as is just about any computer of the last decade+. The CPUs are beasts, the disks are screaming fast.

The operating system itself may not be ideal at serving but you can just run Docker/Orbstack if you need to do something especially Linux-y.

I’d put the question back on you — what are scenarios where an Apple Silicon Mac wouldn’t cut it as a light server for one person or a handful of people? About the only scenario that comes to mind is scenarios where you expect to utilize it so heavily that the fans are running for many hours a day. At some point those are either gonna wear out or just ingest so much dust that the machine runs hotter and needs a deep clean. But even that is largely mitigated by just pointing an external fan at it.

My Mac mini m1 is a fantastic home server.

Years ago I realized running a 100W PC all the time was REALLY expensive, so I switched to an old linux laptop at 30W (about $4/month). That helped, but I moved to the m1 mac specifically for the power efficiency. It does all the things my linux server did, and it does them at 6W (75¢/month).

I've found no competitor with similar performance that can run in such a low power footprint. M4 mini's are way better at power/performance, and I imagine their price is about to come way down since the M6 mini just got announced.

Software-wise, it's different, but mostly equivalent. Homebrew or macports has a similar software inventory to debian. And Apple's container framework is a welcome improvement over colima for running most container workloads.

That too, it's either Mac mini or Rpi if you care about power, and the Rpi is very limited.

My home server is an ASUS NUC 14 Essential which idles at 5W with a bunch of services running. Its peak performance is probably less than the Mac Mini, but the hardware is cheap (at least it was before RAM/storage got expensive) and it’s a great platform to run Linux on.

NUC is good too. Having x86 is surprisingly important in Linux.

It takes almost 8 years for $3.25/mo of electricity savings to break even with a $300 difference in up front cost. Assuming the time value of money is zero, that is.

Yeah, the reality is you never break even. It's only a thing if the two options are similar pricing.

I'd never go x86 again after owning an m1. I'd have replaced an x86 laptop 2 or 3 times by now (2020 m1). My last $3500 dell xps 13z before buying the m1 was absolutely horrible.

I can't wait for an actual competitor to the M chips from apple. It's frustrating.

What CPU / GPU combination would you recommend? Can use unified host+device memory?

For a desktop you may find yourself better on a Linux or Windows machine price/performance wise.

I personally own an M3 ultra, an M1 max as laptops, but my desktop is a Ryzen desktop I built in 2022 and it was a third in price of the ultra for more power.

Recently took a Minisforum 7840hs PC out of rotation as a media PC and made it a full time coding workstation with Proxmox. I do a VM per project due to the nature of agentic editors.

I was using a VM setup on my MBP but it felt like a huge waste, having to leave a laptop on 24/7 when all it did was run Claude Code inside VMs.

I likely will stick with a Macbook Air 15" for next purchase, and beef up my "Claude Server" down the road.

Yep, this is my exact thought. The pendulum has swung back toward a desktop making more sense for me than a laptop. It all depends on whether there is anything useful to do with an amount of computation that can't be fit into a laptop package. For a long time there wasn't, now there is.

This is my setup (except I have an Air because the Neo didn't exist yet). And it's great. Tailscale makes it trivial.

So, do you use VNC (or ssh) to access the Mac Studio at home? Or do you only access other stuff that's living in your home network?

not OP but tailscale ssh does the trick for me.

ssh in terminal and/or through VS Code. When needed, I also use Apple's Screen Sharing app.

>I do wonder if my next computer should be a Mac Studio instead of a MBP that lives its life docked.

Having owned 3 MacBook Pros since 2008, the decision to make my next computer be a Mac Studio came down to (1) MacBook thermal throttling that slows down CPUs when it starts to overheat and (2) easier upgrade of Mac Studio SSD with after-market storage module whereas the MacBook requires more complicated disassembly and hot air gun to dislodge the surface mounted SSDs.

I have a brand new M5 Pro MacBook Pro I don't like it when the fans turn on. The Mac Studio will be faster and quieter for the same workloads.

You don't want a computer that does thermal throttling but also do not like it when the fans turn on?

>but also do not like it when the fans turn on?

Sorry for not being clearer. I don't like the MacBook's noise when the fans turn on.

The Mac Studio has bigger heat sinks to delay the need for thermal management -- and if its fan does need to turn on, the bigger size means it's still silent instead of the high-pitched whooshing noise the tiny fans make in the MacBook.

You can get large under-laptop cooling pads with giant fans that can help this somewhat. It's not part of the laptop, but help if it's docked at home. It's not totally ideal since the bottom of your laptop doesn't have radiator fins, but does help.

I used to game on my 2014 MBP. That was a bad idea - took about 5 years, but eventually my batteries became spicy pillows.

Turns out that even with the fans at full blast, the batteries didn’t like being so hot over prolonged periods.

I’m happy with my M4 mini now, with a separate windows pc for gaming.

I have been thinking a lot about buying a very fast desktop and a cheap laptop that uses remote desktop to connect to the fast computer.

But in the end I ended op buying a Lenovo Legion and put Linux on it.

Laptops are so fast these days that I didn't want to be bothered with setting up connectivity to a remote desktop.

But if your laptop never leaves your desk I think a desktop computer is a great option. Relatively cheaper and easier to maintain and upgrade.

At my current job, one of my biggest blunders was thinking "Let me just order the same hardware as most of my teammates to avoid unnecessary complications".

Most, if not all, of our current work happens on remote cloud vms. Now I'm stuck with carrying a 3KG monstrosity to work.Every day.

Absolutely no positives compared to my ThinkPad that weighed less than half in my previous job.

Remote development is "good enough" these days. With VS Code, Development containers etc... Having a light weight, portable laptop is so much nicer than a laptop that you can't even rest on your lap for long duration.

The only thing you need to be mindful of using light weight laptops is not having enough RAM to fit all your browser tabs..

I totally agree I use containers through propelcode.app which means I can code on any device including my phone and the environment is the same as the server target. It’s even more important with agents that can delete files and change OS settings. I never need to worry that a coding session ends up breaking my main OS environment by deleting or changing a file.

Yeah I stepped down to a smaller MBP. the only reason I didn't go for an air is my home setup with dual monitors involves using the HDMI port. There probably is a dock or something out there though.

I’m doing essentially this, and got a MacBook Neo.

Fir kinda the first time in my life I don’t really have development tools on my personal laptop. I ghostty and openvpn client installed.

I have a large remote linux workstation (2x 8c/16t xeon cpus, 256gb ram, 2x8tb spinning rust disk) and i have my tools over there (along with some VMs).

It works surprisingly well.

Also, the macbook neo is a surprisingly capable little machine.

Apple prices have always been insane, there is a reason why during the days it almost went bankrupt, in Europe it could not rival with PC, Amiga, Atari, Acorn.

Another option is a 'maxed out' mini e.g M5 Pro/64GB. Mini is super easy to travel with if you know there'll be a 'dock' at the far end e.g office<>home or whatever.

So I can somewhat speak to this, as I actually did this for a while c. 2021 and 2022.

The issue (apart from my Mac mini still having the old, bigger form factor back then) is that this requires a full shutdown (obviously), and that is more friction than opening a closed laptop lid. It just takes a moment for every app and background service etc. to settle back in after, and depending on how your brain works, you might not like that a whole lot.

It absolutely is a cool feeling to carry a pretty mighty desktop in the backpack, though.

I think it's actually good to have a forced reboot from time to time, but then I'm not very disciplined in closing stuff I'm no-longer using! Especially like docker containers I was using months ago, that I forgot to shutdown.

Yeah, I agree (with the same unfortunate habit) – but up to twice a day of forced work full stop is overdoing it, too.

Is there no hibernation option ?

You don't get the option in the UI: there's just "Sleep", which on a desktop Mac seems to suspend to RAM only. At least, it seems to on mine, as I found out when I unplugged it.

(Apparently the laptops do a suspend to disk depending on battery remaining, so the OS does support it. It'd be nice to have it available! The disk space required could become annoying if you only bought the 1 TB SSD for your fancy 512 GB Mac though.)

Pretty sure hibernation is dead. Deep sleep states use such nominal power, I think they decided it wasn't worth it.

What if you have some battery system working? Then technically, you could travel with it

And imagine plugging a small iPad screen into the thing with sidecar… Almost like a Macbook!

At that point you lose too much convenience. Maybe if someone built a mini UPS with at least 10000Ah capacity, USB-C 4 pass-through and the right form factor to permanently attach it on top of the mini.

https://www.youtube.com/watch?v=8okiCCZk7U8

Oh! The fact that touch and pencil is natively supported for macos is neat!

I did this for a few years, going further back in time.

The shutdown thing didn't bother me, and in fact the Mini was lighter than a Macbook Air at the time. So all in all it was pretty convenient.

> that this requires a full shutdown

The mini sips a truly tiny amount of power. I wonder what the smallest/lightest UPS could could bundle it with would be.

But at this point.. you’ve got a shit laptop.

Does macOS not have hibernate/suspend-to-disk?

A part of my brain can’t process acquiring a Mac Mini M5 Pro knowing that they sell M6 despite being aware that it is just a weaker processor.

It will get worse with the MacBook ultra rocking a M5 too.

Apple’s marketing department is gonna kill me.

Think about it maybe like:

I want a 2026 pickup truck. The dealer also sells a 2027 hatchback.

I am not shopping for a hatchback.

I had the same view recently. For the past year I’ve been using my iPad to remote into both my MacBook Pro M1 Max and my racked Linux workstation at home using Jump and moonlight/sunshine, respectively. Never looked back

My goto. Even ignoring speed, it's nice to have a remote machine that keeps doing whatever it's doing while your laptop is closed. And the ARM Macs idle at such low power that it's not wasteful like my MacPro4,1 was haha. My UPS's ammeter doesn't even display the Mac mini's draw.

They gave me a nice MBP for my new job. I tried doing heavy work on it locally, it was fine for that, and yet it still ended up being a light terminal into an EC2 instance, partially because their stuff is on AWS and latency is way lower within that. My personal mini is running too.

Yes, except that the only way to get a fancy screen is also to put the fancy cpu in it. If there would be an Air with the screen of the pro I would buy it. Maybe even a Neo with that screen.

Oh yeah. Throw it behind tailscale. Herdr multiplexer in terminal is cool, it’ll remember your session if you open it in ssh from the neo

I would not use Neo, Air with 24 GB ram should be your minimum because you will want to do some work locally even when you're running a VPN/SSH thin client setup.

I'm currently SSH-ing into my workstation from my M4 Air with 24 gb ram and it's ideal for this flow. Slack/editors/clients/browsers/etc. easily gobble up over 16 GB. I have no more dev tools/compilers/source on my client machines, everything is dockered up on remote workstation in isolated VMs (too much supply-chaining).

My only downside to using a pro is not having 120/4k HDMI port on air and my dock won't support it with Apple (it does with windows)

Can that run qwen? Also can you explain what you meant by the downside ?

If I was going to run LLMs locally I would definitely go for a pro.

I meant the only downside of not using a pro for my use-case is that I don't have a HDMI port on device with high refresh rate. I have a decent dock that can give me 4k/60hz HDMI but not 120hz refresh for macs (it works on windows). It's a minor thing but having 120hz is nice when I'm docked at home.

It heavily depends on your workflow and how many (electron) programs you need to run simultaneously. I found that if I’m mindful of closing things I’m not actually using, and let macOS do its swap magic, the Neo will just work.

I’m also the kind of person who close tabs and I like to work on one task at a time.

I'm sorry but that's a bit ridiculous. I'm still using an M1 Pro with 16GB of RAM and I have quite heavy usage - docker, other heavy native apps, a bunch of VS.Codes and coding agents, things running in the terminal, background processes. It's just a big exaggeration that general regular productivity apps "easily gobble up over 16GB"

(Edit: saying this as someone who just paid for a new 128GB M5 Pro Studio -- but that's for LLM's, not regular work)

I have an identical gen mac mini I bought for testing out the apple silicon for my workloads a year and a half ago, except it's 16GB min spec version - I got ram throttled on that thing constantly as soon as I start using it for development. If it was just productivity I could get away with it - but as soon as I start something like JetBrains Rider on it or VSCode with C# language memory pressure is in the yellow and if I start a next server for frontend the machine just grinds to a halt. Even when I'm using remote workstation my memory usage is hovering above 20gb with active swap on my laptop. Soo many electron productivity apps eating up ram like crazy.

If it was more readily available on retailers such as Amazon I would even recommend 32gb ram air for local workflows - if you're doing development these days you're just going to need it - especially with LLM workflows and parallelizing work. The price difference between 16 and 24 is not even close to not justify picking it for a work machine.

Thing is having more cores/faster cores might make your workflows 1-2x faster but hitting swap is going to be like >10x degradation

Cheap Linux laptop with Tailscale back into the Studio when traveling?

No iTerm2, weird shortcuts, random issues in video calls?, also not really cheap to get a laptop that reliably works with Linux and has decent battery life

Im writing this from a MacBook Air.

Everything is a container or VM now, and none of it runs locally for me.

I have a desktop (older intel) with giant monitors and a keyboard for when I sit at the desk. I have the laptop for when I travel, go out or just want to work from the couch.

When I do my next upgrade to "better hardware" I'm not migrating a machine, rather I'm migrating the containers. My workflow is such that if I loose one of the boxes I sit at to a cup of coffee I really wont care other than the financial loss of a new laptop or keyboard.

The biggest win in all this was dumping the off the shelf firewall/router and moving to Opnsense. Wireguard vpn lets me route all my traffic through home for all my devices (and what is now a growing home lab).

There are scopes of work that this setup would not work for. I would not want to be a video editor with this set up, it's not ideal if you want to play AAA games. But for what I do, it is pretty ideal.

I used to have a high-powered laptop, but just use a combination of ssh and NFS to my machine at home from a low powered laptop for travel these days. It works well.

Makes managing both backups and handling failure scenarios involving loss or unauthorised access to the laptop less of a hassle as well.

Why not a non-mac cheaper desktop/workstation and a Mac laptop for use?

E.g. what most major tech companies have - a laptop that is for VSCode via ssh/web browsing, and a beefy Linux dev box you ssh into for everything else?

It's way cheaper, and what I use at home too - a lot cheaper than a Mac studio for everything, especially with RAM and storage.

A lot more upgradable too. Those empty DRAM slots can get populated one day!

The unified memory is the big appeal of a beefy Mac Studio

if you do want to do local LLM stuff, mac studio is significantly better than most linux dev boxes.

As of a year ago, the strix halo AMD (system on chip, unified ram) was within single digit percentages of the mac, but it’s a linux box, so much better for most dev and as an llm server.

I know AMD has announced the replacement, and it will be faster, but I’m not sure when it’ll ship.

Also, you can cluster strix halo if you need more than 128GB of ram.

Better than a DGX Spark running Linux? Absolutely not (I own both).

Real-world matmul is about 4x higher on a DGX Spark than a M3 Ultra.

For specifically unified memory reasons, Apple machines are much, much better at running local LLMs.

I’ve debated Mac Studio or MBP at each generation. For equal spec machines the price premium to get a MBP over the Studio is always smaller than I expected, so I pay the extra amount and get the MBP.

This changes when you get into connotations that aren’t available in the laptop form factor, but with RAM prices the way they are those configurations are more than I want to spend on a local machine right now.

For running local LLMs the high memory Mac options always look appealing, but the processing speed (prefill) is so much slower than GPUs that it hurts. For situations where you have no rush and can let something work in the background for 24 hours it can some times be ignored, but the speeds I get from a real GPU setup are so much faster that I never use the larger models on a high memory Mac any more.

I got an M1 Max studio w/ 64gb when they first came out and I love it. Price was obviously a lot more reasonable at the time. I’m fortunate to have dedicated office space, and there is definitely a nice psychological effect to having a desktop in a specific spot and “going to the office”. I’ve since purchased an m3 MacBook Air as my thin client and for lightweight travel use, and when my wife needs to take a call in the office or whatever I pop out to the living room and either hop on Remote Desktop or SSH into my Studio. Highly recommend this setup. If you’re primarily doing thin client on a local network, a neo should be fine for that (and to use as a media player/communicator around the house). That said, I have been eyeballing Omarchy…

Happy with M4 Mini, although I recommend using a well ventilated M.2 enclosure instead of the integrated docks as I just had a 2TB Kingston die which I'm pretty sure was caused by "zero thermal forks given (chopsticks only)" UGreen hub design. Also can use as a laptop and replace screen/keyboard/whatever when failed/unsuitable. Higher resulting device longevity, especially with Asahi Linux making such great progress. https://github.com/vk2diy/hackbook-m4-mini

I got an OWC Express 1M2 precisely because of thermal fears. The heatsink is huge, but no fans. Only real issue is needing to be careful to not nudge the cables.

I also suspect that trying to stream larger LLM models from disk hammered the USB 4 connection too much - lead to system hangs, so now stream from the internal SSD instead.

Cool product, thanks. Sent back for Kingston RMA, expecting a new drive soon.

I’ve wanted the idea of a home server or even home “mainframe” and some terminals for me and family forever but the idea never takes off….

Same. For me, it’s one of those “solutions in search of a problem.”

It just seems somewhat more efficient to take all the computing storage, memory processes, processors, or at least all the money that go into it… the in one spot.

If you’re running a big task or small, everything scales to the appropriate size, regardless of the hardware sitting on your lap or under your desk.

At least that’s how I think of it.

I've been using the free version of TailScale to use macOS's built-in screen sharing app to use my home Mac mini from my work MacBook Pro. Works great! That combined with ChatGPT/Codex's "remote" feature is an amazing combination. I think you'd be totally fine with a Mac Studio + Neo.

I have been considering this.

I bought a Macbook Air in the interm waiting to see what the Macbook Ultra looks like, but honestly—it's the best form factor ever. I love it! Even though it's only like a pound heavier, the Pro feels like a monster.

Thinking about getting a Mini or Studio with more horsepower to stay at home.

Any other apps or suggestions? I'm still not exactly sure what working this way looks like.

I've been running a desktop Mac in addition to MacBook and iPad for the past decade (currently a first gen Studio still holding quite strong). It's primarily a focus thing, secondarily it's a way to keep work happening locally while I'm elsewhere, and never have the frustrations of docking and screen geometries all re-arranging, etc. The iPad is at the opposite extreme, it's for recreation, burn out management. The MacBook exists for when I actually need to do "real work" outside of the house.

YMMV, but when I got my iPad I realized that my "mobile" needs were only contempt consumption and my MBP was used as a desktop computer connected to a 32'' monitor. So every time I renewed my laptop the expensive screen went with it, and I had to pay for another one that I would be using as a second screen!

Thus, my current workhorse is a Studio.

I almost bought a studio a couple years ago, but I was starting a PhD and so I figured I had to prioritize MBP.

These prices make that decision so bittersweet. I feel so shut out from being able to use AI as economic productivity without being able to take on debt for a local-inf capable mac studio

Last year I was able to justify a 128GB/1TB M4 Max Studio for 4600€. This cycle's M5 Max with 128GB/1TB goes for 6219€. The M5 has to be obviously a better machine, but it's now above the price point I would pull the trigger for. Guess I got lucky in the futureproofing lottery this time, anyway.

> I realized that my "mobile" needs were only contempt consumption

Is that the new phrase for social media? I hope it is...

My contempt consumption is reading feeds, books and sheet music, and some casual games...

... and a lot more social media I'd like to admit.

> not that Apple prices weren’t insane before the ram/ssd shortages

Funny that after the price started to increase last year (I think October/November), there was a window of a few months where Apple prices stayed the same as they were before, thus making apple prices actually good when compared to the rest of the market.

It was a unique opportunity to have acquired a 512G M3 ultra for $10k.

not as crazy a config, but I got a 128GB M5max the day before the price increase for $2K less. still proud of the call to buy it.

Also proud of having spent $2.5k on a used 128G M1 Ultra back in September 2024.

Despite being outdated in terms of compute, it stills let me run very good recent models locally, with Deepseek V4 Flash 0731 being the greatest one right now, and hopefully Qwen 3.8 Flash will also fit well when it is released tomorrow!

The M1 ultra definitely leaves to be desired in terms of its token speeds, but I think 20 tps generation and ~200 tps prompt processing (which is what I get with DSv4 flash), is already enough to do a lot of serious work when you combine with the decent prompt caching provided by llama.cpp.

You're killing me here lol

I have a Mac Studio M2 Max as my main driver. It's fantastic. I will get the best M5 Studio I can afford when the higher RAM options come out because I am very involved with local inference. I find that with Tailscale I can either use screen sharing if I've got a good enough internet connection or just SSH and Mosh if it's a little bit laggy. That has been good enough for me. A Mac Studio and a Neo as a companion is a really good combination. My laptop is an Apple M4 Air but it basically does nothing except for running ScreenShare and a terminal.

> I do wonder if my next computer should be a Mac Studio

Would you get away with a Mini?

Slightly off-topic: I decided to go with a Mac Mini to start playing with AI. My daily driver is a Linux laptop (with a negligible GPU). I also have a few Raspberry Pi machines for backup, and a low-end AWS server.

I was always cobbling together ad hoc network access, to get from one machine to another. Access to my Mac Mini while traveling was a pain. Bringing it with me is ridiculous, and network access was a PITA.

Tailscale is wonderful magic. Free (for my usage), so easy to set up, and now no matter where my various computers are, they are all accessible trivially via a single ssh connection.

So get a beefed up desktop Mac, set up Tailscale, and then use any random laptop, anywhere, to use it headless.

Yeah, I have an MBP for Reasons, but it sits docked 90% of the time. I'd rather spend the engineering budget on better thermals and more ports; I always have it plugged into a fixed monitor and other peripherals.

I’m in the same camp as you. I’ve just decided that the M5 Ultra is gonna be my next machine. Plus an Air for the few times when I need a computer on the go. I’m invested in Tailscale already so an ubiquitous Studio Ultra seems like a sweet idea

I don’t even know what a Mac Ultra looks like and I want one.

It will be new Mac with more GPU cores for ai, right?

It would sit in my basement server room providing local inference.

I have exactly this setup and use a Macbook Air with Tailscale to just connect via High Performance Screen Sharing both at home and on the go (when not sitting at my desk, I move around a lot around the house after 3pm or so while I keep my workflow identical)

HPSS works well via 5G, and Tailscale is incredible for the setup as much as the hardware.

Edit: I ended up getting a 15" M5 Air but honestly trying it on a 13" M4 Air was actually superior because when all you care is mobility (since the Studio does the heavy work) its nice just going around with almost no weight / bulk.

I almost forgot that HPSS was a thing, thanks for the reminder.

https://support.apple.com/en-my/guide/remote-desktop/apdf8e0...

M4 Max Studio + Macbook Air. I've been running this for several years. Mac is a really poor server environment and none of the apple ecosystem makes up for it. So, I still prefer my linux server, but the M4 Max is much more suited for inference tasks so here it is. I wouldn't say it's much better than a MBP except the sustained throughput is much higher and the fan noise is barely noticeable. I'd get a linux equivalent if there was one.

I used laptops for years.

However, after retiring, I realized I never undocked my MBP.

So I got an M4Pro Mini, and I've been thrilled. If I ever get to where I travel a lot, again, I'll get a laptop, but I don't see a need, right now.

I can't really imagine buying a new computer in this market. But I also wonder how long it will be before prices drop again, if ever.

Oh you will be imagining some time soon.

If it's your remote-in laptop, I'd suggest getting a used MacBook Pro or Air with an M series processor in your preferred 13/15" size and memory/storage configuration. You can get a used MacBook Air M3 with 16GB of RAM for about the same price as a MacBook Neo. It'll have similar single core, GPU, and NPU performance as the Neo and about 38% better multi-core performance. Plus you'll have 16GB of RAM for future-proofing and apps. And you can spend a bit less if you want or a bit more depending on what you need. You could get a used MacBook Air M1 8GB for under $400.

I wonder if apple could or would offer a subscription for a OSX VM that you could remote into? That plus a neo would be an excellent combo for me

Yes, and..

Just Tailscale into the Studio from iPad Pro 13" with magic keyboard and Kit Knox's rootshell:

https://github.com/kitknox/rootshell

Note that the iPad Pro can also drive a 4K second screen if you like, and most anything else a Macbook with a single port could drive.

Why not Macbook Air? Because the iPad Pro is also a tablet, touch screen, and 5G...

this is exactly my setup! the ipad pro 11 m5 cellular is such an amazing device especially with the 5G. only downside is obv the iPadOS, it's still very limiting and shortcuts and power tools (like easily prefilling passwords) are non-existent. for passwords, you still have to manually click the pre-fill and/or enable full keyboard to be able to tab into it but still annoying.

thanks for this link! I was constantly struggling with remote-ipad setup - tried so many things (including paid options), but everything was _iffy_ at best (on-call4live, lol)

And I'm also gonna ride my M1Max Studio until it dies (hopefully it will outlive me xD)

Yeah, I was about to say "or an iPad Pro", but the Neo is shockingly cheaper than an iPad Air.

There's definitely an option for everyone if you want to go down the "fixed supercomputer, remote terminal" route.

Why use mac at all? If you want a docked machine you can build a PC with the exact specs you want. Throw a Linux distro on it and you will have a similar experience

Because it's NOT the same experience. iCloud access, shared messages, Cut and paste between other Mac devices..I say this with a Ryzen 9 + rtx4090 super right next to the macs. (A studio and an M1 MBP with a new battery).

While they're theoretically at parity, in actual use, they're not.

This was my perspective for a long time. The issue is the Macbook Pro screen is awesome, and the display story for desktop computers is pretty grim.

10 grand for 256GB memory. Likely double that for 512GB, but won't be available or finalized until October. Thunderbolt 5 is highest bandwidth external IO available at 120Gb/s. 1.2TB/s claimed max internal memory bandwidth.

Not exactly "future proof" for >1T parameter models but good for targeting specific lower-parameter models, or if you can rely on pipeline parallelism and run a cluster.

3 of those thunderbolt 5 ports, so you can do a fully connected 4 machine cluster topology.

It's unclear to me how bandwidth scales with multiple connections. Many-to-many does not seem ideal. Daisy chaining would be fine for straight pipeline work. There doesn't seem to be an equivalent of a ethernet switch for thunderbolt 5 though.

Apple put on the announce page that 4 Studios in this config can run inference at 3x the speed as 1 Studio.

I read each has an independent controller for full bandwidth.

> Not exactly "future proof"

Computers are never "future proof".

> Computers are never "future proof".

Upgradeable components however could go a loooong stretch towards that goal. It can't be that hard to follow a common form factor for at least the housing across two or three generations to allow a reuse of everything but the main PCB.

If it was upgradable, then yes, spending more on top of it every year would make it future proof, but that's not the point. It's that spending 10 grand doesn't get you a future proof computer today.

The way the Apple M-series does ram that might be difficult to pull off.

> The way the Apple M-series does ram that might be difficult to pull off.

Well it might be an idea to keep the layout of the mainboard and connectors the same.

That way, instead of having to upgrade the whole machine, all it would need is a new mainboard. Framework for example managed to pull that off, and in mobile at that, where constraints are much worse than for a desktop computer.

> Framework for example managed to pull that off, and in mobile at that, where constraints are much worse than for a desktop computer.

It's not the same thing though. On the M-series, CPU and GPU share a unified memory architecture and ram is much more tightly coupled to get it to go faster. A closer example would be the Framework desktop, actually, where memory is also soldered in for the same reason.

That actually would be interesting - yes, the computer itself is basically a PCB, but it’s also wrapped in a couple pounds of aluminum, a power supply, cooling fans, and a few other things that don’t need to be consumables. Upgrade a mini or a studio by swapping the new board into the old case - yeah, you’re not saving much money, but you also don’t need to throw out the entire rest of the case, and you can ship the main board in the space of a couple CD cases.

It’s a very non-Apple thing to do, but it’d be pretty awesome if they did.

The main PCB is pretty much everything that has value. The rest is a heatsink, case and PSU.

Think it would have same memory bandwidth if the RAM was upgradeable?

Would be nice if someone knowledgeable about electrical engineering and manufacturing processes could lay out some valid reasons for manufacturers to integrate RAM onto the motherboard.

https://news.ycombinator.com/item?id=49041256#49082206

It isn't integrated into the motherboard, it's integrated onto the same package as the CPU/GPU which allows for better signal integrity and higher speeds. You can get somewhat close to the same speeds while modular with tech like LPCAMM2, but there are some pretty difficult challenges to overcome to close the gap completely. As an example, the Framework Laptop 13 Pro CPUs support up to 9600MT/s (same as M5), and Micron sells LPCAMM2 modules that can run at 8533MT/s, but the 13 Pro only officially supports 7467MT/s.

Each connector will introduce "insertion loss" to the chain, eating into your signal integrity budget.

The SO DIMM Ram slot, which is designed, idk, 30 years ago? is not really capable of handling the frequency we are targeting (close to 10GT/s)

About 20 years ago my dad bought me a $5k computer, it was future proof for about "5 years" before we had to upgrade its internal parts (more memory, new graphics card).

It was future proof but not really because it struggled a lot in its final years.

My 6 year old mediocre gaming PC is still a mediocre gaming PC.

I expect it to stay a mediocre gaming PC for the next 3 years maybe 5 years.

16GB RAM RTX 3060 TI (8 GB VRAM)

Nice. Curious, what do you play with that thing? Do you think you'll still play those in 3-5 years? And are there games you do want to play but can't?

I like to test new releases that are story heavy.

I also like 4K with 90 to 120fps which leads me most of the time to use almost minimal settings.

Look at steam hardware reviews. My setup is quite the default thing. Games optimize for it

He had that option,l. However buying apple hardware todqy you are stuck and need to replace all of it even if for example the CPU is plenty fast but you need more memory or GPU power.

I’d say something like a PS4 is future proof. $350 got like 10 years of modern software.

Nobody calls a PS4 a computer. We all know it is internally with the hardware, but the whole package makes it not a computer.

PS4 is as close as it gets to the colloquial computer; it's a AMD APU running BSD with homebrew linux builds available.

And if not, would you consider a raspberry pi a computer? or your iphone a computer?

i'd say it depends (as with everything)

i'm still using an old i7 3770k @ 4.8ghz with 16gb (ddr3) ram running linux for random tasks like executing tests. obviously, power consumption is higher.

my main machine is a MBP M1 Max which i use for everything. i also have my main linux desktop workstation that has a 5950x with 128gb (ddr4) ram.

i'll probably get 2-5 years out of my MBP, and my AMD workstation will probably be good for another 5-10 years.

i'm not a gamer, but i have a 3080. i'm sure my 5950x will still be good for gaming in 10 years if paired with a modern GPU.

They do however last a much longer time now.

A 8 years old graphics card can still play modern games. Ten years ago playing a modern game on hardware that old would've been unthinkable.

And with how the market is right now, we'll be stuck on the current "reference level" of hardware for a while longer.

I recently got a refurbished Strix Halo 128GB desktop for a steal ($2200) and I have a feeling it's going to be my last PC purchase for the next 3-5 years. Who knows how long this RAMpocalypse will last. My Lunar Lake 32GB laptop is already a couple years old, and it's not a speed demon but it ticks all the laptop boxes (portable, great battery).

>Computers are never "future proof". Till now.

My thinkpad t14g1 with 16gb ram somehow still relevant today.

My main PC at home is 15 years old at this point, running linux. It's fine.

I don't play games, but it still chugs along perfectly fine and for what I use it for I rarely think that it's too slow.

Same thing, I still have at home a desktop I've built when I was 17 (now I'm 30!) and still runs fine for day to day tasks, of course with Linux on it. It has 8Gb of RAM (DDR3), 512Gb SSD (not original of course because back then they did cost an astronomic price) and an i7 CPU that I've overclocked, an NVIDIA GPU with 2Gb of memory that still plays older games fine.

To me there is no reason to update a PC if you don't use Windows or MacOS that forces you to purchase a new hardware to do the same things. Just look at Windows 11, full of useless AI features, weights a ton, and in the end it's probably faster a PC from 25 years ago with Windows XP.

> 10 grand for 256GB memory.

A NVIDIA RTX 6000, 96 GB at 1.7 TB/s, is 13 grand.

This 256 GB at 1.2 TB/s Mac is extremely competitive, it will be sold out everywhere.

Except the RTX 6000 will run circles around the Mac studio in just about every way. Memory bandwidth is literally the only spec where Apple is competitive, and while high memory bandwidth is necessary for LLMs to perform well, many people strangely don't understand that memory bandwidth alone is not sufficient.

It better because you’ll need a few of them to run some larger models (I’ll be just as vague citing which models).

I was not speaking about specific models so I didn’t feel the need to cite any. Not sure why the backhanded insult was necessary. You need multiple Mac Studios to run the largest models as well, so neither is a one-size-fits-all device.

If you can show me a model for which a Mac is faster than the RTX 6000 then I’ll be happy to update or retract my statement.

Energy efficiency is in another league with the Mac Studio for local workloads.

I can run agents using deepseek v4 flash or Qwen 3.8 on my m3 ultra and it will be lukewarm and the fan will eventually start blowing softly.

I’ve run the numbers on this, and an optimized nVidia build can meet or beat Apple platforms in terms of tokens per watt, which is the most important efficiency metric if what you care about is using the least amount of energy to generate a given response.

Yes, the Mac might get lukewarm, but it will take 2-3+ times longer to do the same task.

Yes I agree that at full speed with parallel workloads, the tokens per watt are better using Nvidia GPU servers.

Also while the pre fill performance sucks, the Mac isn’t that slow and can use much better model compared to a similarly priced Nvidia workstation so it’s not really taking much longer in practice.

It’s taking longer than in the cloud for sure. At least a local computer uses the local energy grid that is pretty clean and not fossil energy.

Okay, so you said “Energy efficiency is in another league with the Mac Studio”; I assume you meant that it was more efficient than an nVidia GPU setup. But when pressed, you agreed that’s not actually the case. So it’s not actually that much different in terms of energy efficiency.

> Also while the pre fill performance sucks, the Mac isn’t that slow and can use much better model compared to a similarly priced Nvidia workstation so it’s not really taking much longer in practice.

This is exactly backwards; assuming you mean that the Mac has more RAM so you can use a larger model, the mac is going to be _even slower_ since the fastest Mac isn’t as fast as the average nVidia setup.

My point is that comparing apples to apples (no pun intended), an nVidia setup is both faster and (possibly with some tweaking) more efficient than a Mac.

Mac studio wins in memory capacity, price, perf/watt and value.

RTX 6000 wins in performance, if your model can fit into the VRAM.

There are very obvious and clear advantages to a Mac Studio. It's an entire system for one and you're getting a world class CPU as well.

> Mac studio wins in memory capacity, price, perf/watt and value.

[citation needed]. I have personally specced out and built an nvidia GPU-based machine which after some optimization, handily beat the Mac Studio in terms of tokens/watt for LLM inference with most models. This was in the M2 Ultra era, and I haven't run the numbers for the later generations, but nvidia's cards have gotten faster just as Apple's CPUs/GPUs have, so I would guess that it's still possible to do.

> RTX 6000 wins in performance, if your model can fit into the VRAM.

"if your model can fit into the VRAM" can be true for the Mac as well.

> There are very obvious and clear advantages to a Mac Studio.

There are certain advantages for sure, depending on your use case. They may _seem_ to be obvious, but as evidenced above, I believe that many people overestimate the Mac's superiority on the metrics you cite when comparing a Mac vs. a dedicated GPU for LLM inference.

> "if your model can fit into the VRAM" can be true for the Mac as well.

It is much more likely for your model to fit in large unified memory of a Mac than the smaller more limited memory of a GPU. Even going with two 5090s, you now have to shard your model and that is a PITA.

But it turns out that MoE is the solution both for running models on macs of limited computer power means (not as fast as GPUs), and on multiple GPUs that require sharding the model.

> It is much more likely for your model to fit in large unified memory of a Mac than the smaller more limited memory of a GPU.

I bristle at general statements like this when it obviously depends on the specific Mac and GPU in question. But yes, comparing a maxed out M5 Ultra with an RTX 6000, the Mac has much more memory.

> Even going with two 5090s, you now have to shard your model and that is a PITA.

Every modern tool does this for you automatically. It is absolutely not a pain in the least (e.g. llama.cpp ships with pipeline parallelism enabled by default).

Sharding a dense model using Tensor Parallelism (TP) across dual RTX 5090s has a significantly worse performance penalty over PCIe than sharding an MoE model.

If you are using multiple GPUs, MoE is basically going to be your only workable choice unless you can leverage pipeline parallelism (only half your GPUs can work on a prompt at a time, so you need to process prompts back to back in a pipeline setup, and they better be doing similar things because your vram is limited).

Have you ever actually set up a multi-GPU system for inference? Based on my experience you are drastically overstating the problem. Both tensor and pipeline parallelism (without NVLink) produce a machine which is faster than any Mac on the planet, which is what we’re discussing here. Yes, each has pros and cons, and neither scales perfectly linearly. But it works great regardless.

No, and at $4000+ per 5090, I'm unlikely to have any experience anytime soon.

You can parallelize inference with much cheaper GPUs as well!

I’ve got a 4060 ti 16gb, and I’m thinking about getting another. I previously specced out a cluster using multiple 3090s. At the time, the 3090s were going for $700 on eBay. They’re more than that now, but there’s no need to spend $4k per GPU at all.

  [citation needed].

No need. You can infer the logic with this line I wrote:

  RTX 6000 wins in performance, if your model can fit into the VRAM.

I'm not sure what the controversy is here.

You made claims about performance per watt and other metrics which were completely unsubstantiated and aren't backed up by the line you quoted. That's what I was asking for citations about.

What other metrics?

Do I need to quote your own statement to you, verbatim, again?

> Mac studio wins in memory capacity, price, perf/watt and value.

Mac Studio winning in memory capacity is just a fact. That's why I'm confused.

Yes, I agree. What about the other metrics, including performance per watt which I’ve now mentioned four times and you’ve ignored the previous three times?

https://www.youtube.com/watch?v=nwIZ5VI3Eus

This is a good video to watch.

  Yes, I agree.

Glad you agree. I was just confused why you were questioning it. The big biggest advantage for Apple Silicon is that you can get much more VRAM per $ over Nvidia cards. No controversy.

You seem to be acting deliberately obtuse.

> What about [...] performance per watt which I’ve now mentioned ~~four~~ five times and you’ve ignored the previous ~~three~~ four times?

In the video.

No, it's not in the video.

The video was not focused on performance per watt at all, and the best attempt that the video makes at measuring performance per watt actually shows the opposite: that the 5090 system is more efficient than the Mac.

When he's running the 27B model, at about 5:02, he shows that the Mac Studio is pulling 251.5W, and the PC is pulling 315.3W.

Then he shows the results:

Mac: 27.62 tok/s PC: 40.92 tok/s

Doing the math, we arrive at 0.11 tokens/W for the Mac, and 0.13 tokens/W for the PC. The PC is about 20% more efficient.

The only other time he even shows the power usage at all is near the beginning, running a 4B model which is trivially small for both systems.

So once again, I ask, what are the sources for your performance/watt claim?

How's the compute side now, I wonder? Because while the Ultras have impressive memory bandwidth for inference, processing prompts still takes a dog's age on my M3 Ultra. I heard the M5 makes some strides forward in this area, though, and the M7 in particular promises to go a lot further.

M5 is excellent, they’ve finally gotten their own tensor cores.

Good for inference; however if you like to train, data format support and effective performance is limited (M5 Pro). Some hardware features are not exposed or extremely slow.

You’ll be fine for inference, but pales in comparison to what a RTX 6000 Pro can do for compute/matmuls/training.

4x faster prompt processing than M3 Ultra.

The relevant comparison isn't one mac studio to one RTX 6000, it's a 24 channel DDR5 system, which also has ~1.2TB/s of memory bandwidth (or more when Xeon 6 compatible 8800mt/s memory becomes widely available), vastly higher prefill due to more CPU horsepower, orders of magnitude faster networking, can hook into GPU accelerators, can be upgraded etc. A baseline 384GB system from eg Puget is ~30K vs ~12K for the 256GB Mac Studio and you do get value for the money.

So 3x more, plus the cost of a GPU (another 10k?). How is that value for money to get slightly better performance?

It can be more than "slightly", particularly if the model you're interested in (or will be interested in in 6 months) doesn't fit on the mac studio. You also need to account for eg storing 10TB of random checkpoints, load time when experimenting, and so on. When you start actually needing throughput these are all capability gaps in practical use, not just x% benchmark differences.

If you just want to run Qwen 3.8 27B and Deepseek v4 Flash in perpetuity and that's it, there are a lot of solutions that will work and this is a fairly user friendly one.

You can find model variants to scale up to whatever capacity you have. Queen has models that just barely fit in 256gb, I ran them… okay… on my Studio.

A 3x price difference means you can get 3 256GB mac studios which you can connect through thunderbolt and with RDMA a total of 768GB ram with compute/memory bandwidth scaling basically linearly (with a small overhead cost).

Doesn't make any sense since you can chain 3 M5 Ultra studios for the same price to get 768 which is 2x more than your system.

Lastly, I want to clarify that prefill on an x86 CPU is drastically slower than on an M5 Ultra GPU.

Excellent post. Heck yes. And with MRDIMMs coming, we're going to get another >50% boost in throughput per channel real soon, with a massive uptick in max capacity (4x).

It feels like PCIe is a bit of a boat anchor here. There's a SATA->NVMe style transition waiting in the wings to make this all so much better. We really need post-PCIe GPUs. CXL with it's very small low latency flits. This is an "almost certainly not" but I wonder if you could mix PCIe and CXL so you could have the GPU memory expose vmeme as a bunch of CXL.mem pools but still have an otherwise pretty normal GPU. It seems madness that UALink went all in on GPU-to-GPU with no affordances for connecting to host computers.

Prefill is gated by GPU compute, not CPU compute. A Epyc/Xeon CPU will never beat an M5 GPU in prefill. If you want something that can match M5 Ultra in prefill, you'd need to add beefy Nvidia GPU. However, the problem is that the GPU memory is separate from the 384GB system memory. And here lies that advantage of Apple Silicon. It's unified memory and accessible by the CPU and GPU.

I've been using Macs as my primary machine for over three decades – but in the past year a Microcenter-Prebuilt (with a 5070Ti) has helped make Ubuntu my preferable computing environment (I have three Silicon Macs with similar access to 16GB vRAM – the 5070Ti "smokes" them in compute).

But my next machine for LLMs will probably be the Mac Mini (with 64GB vRAM access) for low-$2300s – which is what I paid for my MC-Prebuilt (and am still very happy with – Ubuntu is great).

After decades of installing and forgetting various shades of linux-distro, Ubuntu has kept me "using" the computer as a tool, instead of "just tinkering with it").

10 grand for 256GB new Ultra sounds too cheap in today’s crazy DRAM market, it feels too good to be true.

There is no "future proof" for >1T param models, there is no present or future where you can run a model that size on consumer hardware.

6 Mac Studios for 100k USD. Consider it a rule of thumb now - 100k to run 1T params, scales linearly.

Out of curiosity, what do you do to distribute the weights and inference across the multiple machines?

Thunderbolt 5, and software

at a 100k$ you're not buying apple products to be limited by the 120gb/s thunderbolt port.

If you have 100k you can just buy a few b200s.

One does not simply "buy a few b200". They come in 8 packs minimally, at >5x that budget.

I don't know that "consumer hardware" is a useful distinction anymore, it's just "what's your budget and what's your speed requirement".

Consumer hardware means:

- 120v input plug

- not rack-mounted

- has a video out port

Ah, a MacBook Professional is Consumer /s

Yes?

Pro?

Want to bet that consumers with enough money buy it?

Apple's use of "Professional" is just a fancy alias for high end or premium, it in no way indicates anything about being "for professionals" (most extreme example: "professional" iPhone models)

Hence why they had to make up the "Studio" brand for the workstation market, because they'd already fully removed any meaning from "Professional"

In computing "Pro" is not the opposite of "Consumer".

Putting aside the fact it is a marketing label, "Pro" usually means "designed for work" while "Consumer" (in this context) means "doesn't need a special environment".

In computing the distinction is primarily noise, power and cooling requirements.

If a computer is designed to use home power and is quiet enough to use without annoying people and doesn't require specialist cooling then it is a consumer device, even if it is used for work.

To me, "consumer item" is when i don't have to contact sale people for price.

That's in the same ballpark as two 128GB AI machines like the Asus GX10 or DGX Spark or Strix Halo. And, it seems very likely to perform better than either of those for inference. And, 256GB brings some pretty good models into play.

But, that doesn't make it a good deal. It just means the Apple tax doesn't apply when stacked up against AI machines and with memory prices being so out of whack. I'm still planning to wait until the RAMpocalypse ends before I buy any more hardware.

Is there anything in the works or planned that suggests the RAMpocalypse will end anytime soon? e.g. new fabs being built, permitted, planned, etc...

Lots, if your time horizon is ~5 years for not buying new hardware, you're good.

Otherwise you'll have to wait to see if the AI circular financing club collapses- if you still have a job, there should be deals to be had...

There are signs that the money faucet is being turned down. Various investments that were announced have been quietly canceled or reduced in scale. I don't think it'll happen soon, but it seems unlikely to be more than a couple years. If I were a betting man, 12-18 months seems right. The AI companies that can make enough money will survive, the ones running on investor cash and debt, won't.

Efficiency is improving, both in hardware and in software and in intelligence density (smaller models can effectively do more of the AI work that needs doing), so I think the pure data center plays will falter. If there isn't some other business attached, they're never going to recoup their investment. Anthropic and OpenAI are buying all the compute they can find right now, but efficiency gains, especially those coming out of Chinese labs where they must be more efficient to compete, will make it less and less of a problem.

I mean, think about the hardware we use for AI. It's basically an accident. GPUs were not designed for AI (though they are becoming more focused on AI). The specialized AI hardware industry is just ramping up.

So, we're still early in the curve for how efficient both the hardware and software can be at performing these tasks, and given the effectiveness of recent very small models (e.g. DeepSeek V4 Flash 0731 and Qwen 3.8 27B), I just don't see a long future for giant data centers built around billions of dollars worth of last years graphics cards. As with the crypto mining operations, at some point, it becomes more expensive to run the hardware than it makes in revenue. And, as with the crypto mining operations, when the money dries up, the hardware hits eBay and prices drop.

There was an article last week that a Chinese RAM manufacturer was planning to add new fabs to be able to ramp up. More or less simultaneously there was also an article on Apple considering switching to using china sourced RAM for China destined devices.

Define "soon"

RAM production is completely sold out for 2027[1] which means the prices are locked in until after then.

It takes about 2 years from the time ground if broken for a new fab to be built and producing RAM.

There were some new fabs announced between February and April this year by both the Korean and Chinese manufactures, so that new capacity might start having an impact in 2028 in the most optimistic scenario.

Samsung says supply will remain tight in 2028[2], and Micron says "tight beyond 2027"

The best hope is that new (Chinese) players overbuild fab capacity and supply outstrips demand. That isn't likely, but perhaps in the late 2028-2029 timeframe could happen.

[1] https://www.techpowerup.com/351344/memory-makers-seal-2027-d...

[2] https://www.tweaktown.com/news/112966/memory-shortages-will-...

[3] https://s25.q4cdn.com/621799436/files/doc_events/2026/06/Q3-...

No one making statements has an incentive to tell the truth though (in fact, they're all incentivised to keep proclaiming the RAMpocalypse will never end).

Producers (Samsung, SK Hynix etc.) will not say "prices expected to drop" or "demand expected to drop" even if it was true because then consumers would start delaying purchased.

The big buyers (OpenAI, hyperscalers etc.) have no incentive to say "supply expected to start opening up" because that would imply their growth trajectory is flattening; also a huge part of their moat now is just deployed RAM.

There are some news about AI/LLM progress rate flattening out: People uses the cheaper model more than better model. Some AI startup in Chinese lose half of value.

The memory bandwidth on the Spark and Strix Halo is a fraction of the M5 Ultra. "Perform better" is probably a huge understatement.

Yes, if I was going to throw away ten grand on a computer that can run models that are much worse than what I can rent from a variety of providers for a few bucks a month, I'd buy the Apple.

Where are you getting DeepSeek for a few bucks a month? I spend around $3/day on it, sometimes $4.

I didn't mention DeepSeek, though with the caching effectiveness in Reasonix, you can get DeepSeek for pretty danged cheap directly from deepseek.com, even at the higher prices. I was referring to the many subscriptions available that will sell you quite a lot of usage for $20/month. Given how slow local inference is, you're probably not going to ever get your money back out of it if you just use it for inference (or training/fine-tuning, since you can rent GPUs for a buck or two an hour and ten grand is a lot of hours).

When memory prices come back down, I'll be down to the Apple Store. But, it doesn't make sense to buy hardware right now.

I don't think the $20/month packages come with API access at all.

I bought a refurbished M3 Max a couple of years ago just to try things out, I get 90 toks/s which is good enough for a lot of tasks. But ya, when DRAM prices come back down I'll definitely look for a better local rig.

The Qwen and Z.ai coding plans include API access, last I checked. OpenCode Go also has API access to a bunch of models for $10/month.

It will always be wasteful to have a GPU sleeping next to you 99% of the time. Something like OpenRouter is the solution imho, for me at least. I realize that that some people care about privacy more though, and I respect that.

Reading this press release was almost physically painful, because of how much Apple loves to use the phrase "up to". In total it appears 46 times.

up to 46 times, you mean?

Up to 47.

That would be oddly precise. More like "up to 50", or depending on speaker, "up to 100". /s

It would also be incorrect. "Up to" doesn't mean any possible upper bound, it describes the top of a known range. I can't say a car I'm selling goes "up to" 900mph if it tops out at 120mph.

Well, depends. Maybe it can go up to 900mph¹?

¹ drivers are responsible to get the car into a low earth orbit themselves and need to be familiar with the orbital maneuvers needed to accelerate the car to 900mph

Whats an up to dog?

A lot of those "up to"s are expressing user options, though -- "up to" 512 GB RAM means that there are configurations that include that, but also cheaper configurations that don't. That's a pretty different beast than "up to 4.3x faster performance", which means (presumably) that exactly one benchmark showed a 4.3x improvement and the rest showed less -- which is to say, means very little.

Except for the prices of course, where they use “starting at 5,499$” instead of “up to 18,299$” to indicate user options.

The cleaning cloth went down in price about 50%

There is some consolation there

Only because it doesn't contain RAM.

Don't give them any ideas

Moore's Cloth

But was the compatibility table updated?

I thought it was funny years ago when apple would report the power in watts of some systems like the mac mini.

But for the mac pro, they would say something very obscure.

"100-120V AC or 200-240V AC (wide-range power supply input voltage)" and the maximum current is "12A (low-voltage range) or 6A (high-voltage range)".

They're trying NOT to say it had a 1440 watt power supply.

> They're trying NOT to say it had a 1440 watt power supply.

It isn't, though. That rating is for circuit sizing. Power supplies are rated by their output power. If you assume an 89% efficiency for the power supply, and assume 1.0 PFC, and assume that it must be able to supply its rated power at the critical corner (low voltage), and assume zero margin on the current rating, then this is a 1068 W power supply. More likely its design spec is around 1000 W; it certainly can't be 1440 W.

Oh, that's for the 256GB option. Based on their math, I'd hazard the 512GB option being $23,799+. ($4K to go from 96 to 256)

> A lot of those

46, actually. 46. I am pretty certain that the explanation lies completely elsewhere...

(overuse of overstatements, for the sake of manipulation, perhaps? something that makes their saying unreliable this way - e.g. they forget to mention the other end of the range, or perhaps the mean, completely .... except when it is good for them, e.g. prices ... that's borderline to lying this way...)

That's probably up to 300% of an increase from 'previous' press releases

I enjoyed "up to an 80-core GPU, and a staggering 512GB of unified memory"

Staggering?? I can count on no hands the number of times that a fact or figure has caused me to stagger.

Idk 512GB of (essentially) ram is pretty damn insane for a consumer machine

This thing is going to cost >$10K, probably, which I think means it's not in the consumer segment

For the 512GB? You're going to be over $20K, fully specced.

Yes, but you've surely met people who giggled when they read that.

When you get that credit card bill for the 80 core, 512 gig, 16 TB storage box, you might stagger a bit.

Thanks for the laugh!

A silly hobby of mine has been to track Apple's use of the word "computer". They don't use it. The only use of the word that I can find, on apple.com, is in legal boilerplate.

I want a computer dammit, not an appliance.

They also don't use "appliance." Nor do their computers behave as appliances.

It's just branding - they don't want to sell you a generic "computer" - they want to sell you a Mac. It's really not all that different from a computer. Might look a little nicer[1]. Uses a custom processor. But you can write software for it like a computer. And it can display things on a monitor like a computer.

They do have wrist, pocket, and tablet versions of computers, but those might be considered more like appliances.

1 - taste being subjective and all

What's a computer?

It appears up to 46 times.

I am sure I'd love computer products more without the nauseating marketing bs. With dry boring technical facts only. I am repelled now, need more self drive and effort getting to the cashier, than without it.

(Scroll-oriented web pages see me running away fast. Not only Apple, which is a champion in it, but articles in general. That make me a passive spectator eating what they want me to eat, the way they want me to. Rather than a free discoverer. I hate being pushed into something. I like freedom.)

> Apple’s most powerful Mac raises the bar for local AI

Wow, "Local AI" mentioned in the subheading above the fold - it's really awesome to see Apple leaning into this use case and I think it will definitely pay off for them going forward. Fingers crossed Apple is able to put some engineering effort towards shipping with one of the frontier open weight models included and optimized exactly for the machine.

The local ai is used by Apple services, mostly. Though there is a Swift SDK to access it, apparently.

You can run Ollama and it’s kind with MLX models for speed.

Tend to be better than Apple AI.

The product marketing managers write for the target audience; of course they know lots of people buying these are the local LLM crowd.

I actually wouldn't want a serious model shipping 'on disk'. Models release so often, it's going to get outdated quickly. LMStudio is trivial to set up.

Now imagine a model shipped in silicon that gives 5,000 tokens per second at Opus 4.8 quality

U2-OSS-vBono-120b

I am absolutely open to being corrected but I can't think of a single current era FOSS project that apple is optimizing and shipping with their computers.

Shipping with their computers -- maybe not, but they did make FoundationDB open source (again) and are developing and maintaining it.

1.2 TB/s bandwidth of M5 Ultra comes from two dies of M5 Max (each 614 GB/s) connected together using 4.4 TB/s inter-die fabric.

For a non-quantized Deepseek V4 flash on an ultra, I would estimate about 1000+ tokens per second prefill and 50+ tokens per second on generation. This is actually quite usable and near parity to cloud.

They mention "adds the GPU Neural Accelerators." which, if exploitable for LLM loads, would probably help the prefill a lot

Yes, and they specifically mention "Up to 10.7x faster LLM prompt processing in LM Studio" which is probably using the neural accelerator for prefill.

How much would is the cost for that machine though, I'm pretty sure I could just buy tokens from a provider and never run out of money for 10 years, and get far better quality output because inference is being served by professionals on far better hardware and this machine would be obsolete long before that as well. Hosting local seems like a possibly the dumbest thing you could possibly do from an economics perspective. And don't hit me with the privacy argument because everyone saying they care about privacy uses fucking gmail, whatsapp and instagram all day long.

Ok sure. But this machine you can haul in an Ice Road Truck to the North Pole and do inference there in an off grid shack. Good luck talking to cloud AI there.

Yeah, that's who Apple is making these for fucking Santa and his Elves lmfao

Wow even the elves will get replaced by AI

You can’t run multiple workers 24/7 for 100 bucks a month

The problem is when you don't want to send your data to any provider. Private documents, health information, all that is something you don't want to send anywhere, even if the provider makes a pinky promise that they won't use your data for training or store it.

And no, I do not use gmail, whatsapp or instagram :-)

you might want to do that, but you underestimate how much some hobbyists/makers/hackers value being able to do things on their own tools in their own house.

for me, it’s significantly more about the process than it is about penny pinching. it’s why im not an accountant and why i often despise bean counters.

my dad has soooo much money in tools alone for working on cars, faaar more than he’d ever spend at a mechanics. add in the costs he spends buying hobby cars and real estate/upkeep for the garage to work on his cars and it adds up quick. people spend a lot on their hobbies.

look at how much woodworkers spend on tools. look at how much homelabbers spend. photographers, boat/sailing wonks, horse owners, and on and on and on.

add in the above, then the value of the privacy aspects, idealism aspects, potential career knowledge, etc… and it’s obvious that local models will do well.

Why is the privacy argument invalid?

There is definitely classes of data that you legally can't share over "fucking gmail, whatsapp and instagram" and you can't send to commodity token providers either.

Taking liberties to save money at the expense of regulatory duties is a good way to get hit with a massive fine if you work in financial or health services.

This is a product aimed at niche professionals doing paid work. If you don't need that just move on to something else.

One way to gauge how close to viable local inference is the quality of the arguments against it are.

$100/mo for ten years is $12K. $200/mo is $24K. When you look at the valuations of the companies providing these tokens, you don't necessarily see these rates going down.

Your argument against privacy is so obviously bad I don't think it needs a response.

Today, most companies and individuals (including me) are still going to be buying tokens from the cloud. But for many it makes sense to go local today, and it seems pretty clear the day where we turn the corner isn't that far away. E.g., suppose the RAMpocalyse eases in 3 years and Apple releases the M7 ultra, which has actually been designed for local inference (and other players are making their moves as well). I suspect it will become something of a no-brainer to do most inference locally, perhaps supplemented by lighter usage of per-token plans from the cloud for access to specialized models.

The thing that intrigues me the most is this:

"Storage performance is up to twice as fast, with a next-generation SSD architecture built on PCIe Gen 6..."

This is the first personal computer I've noticed that has PCIe Gen 6 storage. I've only seen enterprise PCIe Gen 6 SSDs up until now. Gen 5 SSDs in consumer devices already have high temperatures and thermal throttling, so I'm worried about how Apple's implementation will perform (I know they don't use off-the-shelf SSDs anymore, but I'd imagine the temps would still be a problem).

Im curious, how often is SSD / storage speed performance really useful? I feel like for many people it's akin to gigabyte wifi in that its nice to have, but not really particularly necessary

Boot time, loading game time, mostly.

It's nice when using large sample libraries so you can stream them off the drive without as much caching in memory. Local LLM workflows probably also benefit from being able to run models too big for RAM off the drive.

I bet this is a lot of what is driving it on the DC side, and just in general bringing disk closer to RAM as far more conventional database type workloads.

Apple’s “LLM in a flash” paper [1] sheds some light on why this can be a big deal, especially if they are working on codesign of model and hardware in this dimension.

[1] https://arxiv.org/abs/2312.11514

It's very important. Every time you launch an app or open a file it's critical for performance. Also when memory runs out and the OS swaps to disk it makes a huge difference.

it's very very important, not at all something you can compare to WiFi.

data processing, LLMs, model loading, MoE loading, etc, etc relies on very fast storage to keep your GPU saturated.

When I was booting from a spinning disk, at that time it took over a minute to boot. I upgraded the system to SSD later, and that same system took less than 20s. That was 10 years ago.

Another example as a developer, one particular large project took an hour to build on a spinning disk, on a SSD it takes 2mins

So, yeah, it's really noticable improvement

The big deal of a decade ago about SSDs wasn't throughput, but random read/writes. HDDs can be astonishingly fast, but by their very mechanical nature, they suck at jumping from sectors to sectors to gather small files, which is typical of booting an OS.

Unrelated to disk, I've definitely been thinking about this more on the network side recently. I upgraded to 1.4/900 home fibre a while back and then got a modern Unifi UCG/U6 wifi setup, and while I do get 1.4GB down from my hardwired workstation in an explicit speed test I have yet to get anything close to that on real-world workflows. Torrents, container-pulls, backups, all of it seem to cap out around 20-50 MB/s.

So while have having this massive almost-symmetric fibre pipe is cool on paper, I haven't felt a huge need to install 2.5G gear all over the house.

It's definitely important for editing video and doing VFX work, and lots of the editors/VFX folks I know are using Studios. Raw footage from cine cameras now can be 50-100 MB per second now that we have 12 or 18K cameras, and even the low-quality proxies people work with are still pretty heavy, so in-memory caches aren't really an option and stuff is constantly being read from (and cached to) disk.

Should make a big difference when you're running your system off swap cause RAM is too expensive :P

Apple puts the NAND flash controller on package, so it cooled as part of the SoC/CPU as they share an IHS. Given it is part of the fabric it only has to maintain signal coherence on millimeter to micrometer scales, while enterprise/consumer NAND PCIe have to adhere to standards which force them to signal at 800-1200mV and keep that signal stable for centimeter scale distances.

Basically, Apple gets to cheat because they shove everyone onto the same IC & make the OS that runs on this SOC.

They wound up with this design because these chips are descendent from the iPhone 4, where the whole motherboard was 1 inch.

Vertical integration is unbelievably effective.

I just barely got used to gen 5 and haven't even heard of gen 6 yet. Gen4 should be enough for most people these days, but it's good to see progress.

I think this is wrong - it will likely be PCIe gen 5. Apple doesn't say PCIe gen 6 anywhere and this doubling would be from the previous Mac Studio, which probably had PCIe gen 4. So this would match what happened on macbooks recently - upgrade to PCIe gen 5.

I think someone made incorrect conclusions and everyone just copied them.

Apple did say PCIe Gen 6 - in the original version of the press release. That's where the quote came from. But they have now changed it to exclude the reference to Gen 6:

"Storage performance is up to twice as fast, with a next-generation SSD architecture, delivering industry-leading read and write speeds for rapid project loading, file transfers, and loading huge LLMs."

Here's the original, archived press release from August 25th (look under "Blazing-Fast Storage and Pro Connectivity"):

https://web.archive.org/web/20260825130329/https://www.apple...

And here's the first modified version that the Wayback Machine has, from August 26th:

https://web.archive.org/web/20260826020313/https://www.apple...

So it seems like Apple made a mistake in this case. I agree with you that it will likely be PCIe Gen 5 - that's much more logical.

oh wow, I only saw the later version. Such a mistake!

I wish they were offering 1TB of Unified Memory for the M5 Ultra. I already have an M5 Max MBP w/ 128GB of RAM for running local models, and while there's a /few/ models that I can run in 512GB that I can't run in 128GB that are interesting, where things really shift is at 1TB of memory which allows you run >1T parameter models w/ 4 bit quants reliably. 512GB is just on the edge of "enough", which is maybe the point of maximum frustration considering current memory prices.

Personally, I can't justify dropping the dosh for a 512GB M5 Ultra, but I would be able to justify it to myself if I could get 1TB of memory, because it'd guarantee the flexibility with local models I currently am missing. Seems a huge miss to not offer this... for a price.

1tb would likely be ~$20k - given the current >$10k price tag of 256gb. Would you still be considering it at that price?

More likely double that, even. I think you'd still see many buyers there. You can spend like $16k alone on a RTX 6000 PRO with a mere 96GB of VRAM now..

I'd consider a 1tb machine at 20k, but I'm not going to pick up a 256gb one at all. 1TB fits a frontier-ish model in memory without massive quantization, which is a very interesting capability for a non-rack piece of compute.

But why...?

At that point just rent proper GPUs in the cloud, you'd have way more power and pay only what you use for.

Because I spent $15k in AI costs last month doing real work at my day job; I'd like to do the same thing for myself but I don't like lighting cash on fire.

I think that within the next year or two we'll have the ability to put a frontier-class model on our desks, and I like the idea of playing around with that today to get ahead of it. I could be wrong, I often am.

I'm not optimizing for $$$ spent today and agree that this isn't the cheapest solution to a specific capability problem, but I don't super mind paying for a cool piece of compute that I can do some fun (and educational) stuff with.

I would probably spend up to $30k if I could get 1TB of Unified Memory, because it would allow me a guarantee to run pretty much any local model I want, including >1T parameter models with reasonable quants. I wouldn't be surprised if 512GB is close to $20k when it becomes orderable in October. The justification is less about absolute price and more about price to what it enables. 512GB really doesn't enable much over 128GB for me, but 1TB would massively change things.

I have a bit of paranoia/anxiety about AI, but it's not what most people are concerned with. I understand the limits of these tools very well, and still find them extremely useful. What concerns me is that it's going to become difficult to impossible in the future to run local models which have near-SOTA capabilities in a way in which you can exercise full control of the model. I see the writing on the wall, and its more than worth it for me to invest early to ensure my own capabilities. I am very much not a fan of our "you'll own nothing and be happy" directionality for the world, and I am (at least currently) privileged to have the means to slow that decline for my own self.

I suspect we will see more companies which will burn the weights right in to silicon arise. There is already at least one company out there that showed its possible so others will follow IMO. Basically you will see the rise of disposable weights like Nintendo cartridges back in the day. Use it for a time until the better model comes out and you get a new "chip". Though there's a caveat for this business model and that requires you to pump out lots of these chips on the cheap so you are beholden to the lithography companies and what they can produce for you. If you can do this at scale and doesn't require the latest state of the art nm architecture design you are golden...

> There is already at least one company out there that showed its possible so others will follow IMO.

Are your referring to Taalas / chatjimmy ?

I definitely think we'll see an ASIC-like approach in the future, especially for embedded small models where it may require minimal silicon area and can result in near-realtime performance. But at the frontier, I don't think this is a solved problem and will continue to have model weight churn that will advantage more flexible general-purpose hardware.

chaining an option?

RDMA is buggy and Thunderbolt only delivers 1/10th the throughput of native connectivity. 1TB of Unified Memory w/ 1.2TB/s of bandwidth with marginally ~$30k cost is a different story than 1TB of sorta Unified Memory w/ an effective 120GB/s of bandwidth with a marginally ~$40k cost + all the RDMA bugs.

You need latency for token parallelism, not bandwidth. Hence actual RDMA that bypasses the software TCP stack (ROCe or whatever).

They say you can cluster up to four with a shared memory pool, and get three times the inference performance of a single machine.

It looks like speculation that Apple would raise the base chip’s maximum RAM from 32GB to 48GB was wrong.

Apple also launched the base M6 today with a 32GB RAM limit, suggesting 512GB may remain the maximum for Ultra chips for some time. Since these Ultra chips combine 16 base chips:

32GB × 16 = 512GB

They probably literally don't have enough NAND to go around. 768GB of memory (48GB x 16) is enough for nearly 100 iPhone 17s; that's $80k (corrected) of iPhones at MSRP, although likely much lower margins than these high-RAM boxes.

They also cancelled availability of the 512 M3 Ultra months ago in many regions, likely just redirecting memory

800 USD/iPhone x 100 iPhones = 80k USD

Oops, thank you, corrected.

they`re skipping m6 pro and ultra

AFAIK these "rules" are artificial and there's nothing stopping Apple from having twice as much RAM. They just don't want to.

Mac Studio with M5 Max starts at $2,499 (U.S.) and $2,299 (U.S.) for education. Additional configure-to-order options are available at apple.com/mac-studio. Mac Studio with M5 Ultra starts at $5,499 (U.S.) and $5,099 (U.S.).

I am in Europe, and the Mac Studio M5 Ultra GPU 64 cores with 96GB RAM is up to 6.649,00 €. Ouch.

Don't forget that US prices usually do not include the VAT, while EU prices usually do include respective VAT.

A lot of us in America have no sales tax (if that’s what you mean by VAT) and those that do have it at a fraction of EU VAT rates.

> A lot of us in America have no sales tax

Not that many of us, actually. Only Montana, New Hampshire, Oregon, and some parts of Delaware and Alaska have no sales tax. https://commons.wikimedia.org/wiki/File:Sales_tax_by_county....

I'm sure plenty of Idahoans and Washingtonians, for instance, take advantage of their neighbor status for high ticket items like this...

Except that you then have to pay a "use tax" of 6%:

https://tax.idaho.gov/taxes/sales-use/use-tax/online-guide/

This is difficult to enforce, but it seems that "taking advantage" is strictly speaking just tax fraud.

To put this into a perspective, Google helpfully reminds:

> A fully configured IBM Personal Computer AT (Model 5170) with expanded memory and storage cost around $5,795 to $6,000 at its launch in August 1984, which equals roughly $18,600 to $19,300 in 2026 USD.

It would be interesting to compare the costs of a top of the line machine every decade or so. Costs were steadily decreasing until recently.

They still are, if what you want is roughly the same as the prior generations capability with some uplift (making then number up, but say 20% faster or more ram or whatever).

What is changing is that there genuine demand for more capabilities disproportionate to the cost decrease curve. Fab demand and supply constraints have slowed or even reversed some cost decreases - but that is still getting absorbed by the overall systems costs when you are looking at things like laptops. If you all you want is the last decades demand to browse the web and use office - things are cheaper than ever.

John Dvorak said many, many decades ago (80s/90s) that the computer you want will always cost $3000. That statement has been more/less true for some time periods than others, but with some wiggle room I’ve found it to be accurate enough.

Care to guess the approximate price of the MBP I bought earlier this year?

I bought my refurbished M3 Max MBP with 64 GB for $3k a couple of years ago. Before Ethan I never spent more than $2k for a computer.

The “spend” and “want” number might differ, depending on one’s financial state. I know I’ve purchased plenty for less than $3K. But the one I wanted…

I'm old and established (after spending most of my life not), but ya, I get what you are saying. It was a huge leap for me to spend $3k on a laptop.

Before local LLMs became a thing, I had completely lost interest in buying anything but the cheapest laptop. It felt like "personal computing" was a solved problem. But...it actually isn't, and this is exciting.

My heavily-upgraded M1 Max came in slightly over that when I got it five years ago. (Still going very, very strong.)

This new Studio? Can't find a config under $5k I'd bother with. But for the MBPs that number still mostly tracks for the average Pro user. (I buy large and run it into the ground so long I mistake the ground for the computer's remains.)

I buy large and run it into the ground so long I mistake the ground for the computer's remains.

My still-being-used 2012 MBP (which cost me about $3K) says, “hi”.

And, as you point out, the new computers I want blow Dvorak’s hypothesis out of the water. Never would I have guessed 30 years ago that Dvorak would be wrong the other direction on price.

I got a 64GB M1 Max four years ago or so, used. Felt like a ridiculous overspend at the time, but I'm sure smug about it now.

> I am in Europe, and the Mac Studio M5 Ultra GPU 64 cores with 96GB RAM is up to 6.649,00 €. Ouch.

It would be significantly cheaper to fly to a tariff-free country and buy there.

that's what I'm planning to do.

Where?

Technically, California and other state have laws that say you must report this purchase and pay the tax the moment it enters the state.

Sorry for being lazy, but is there a rough breakdown like "You get sonnet level for M5 and Opus for M5 pro, etc.", or is it still speculative. Or put simpler, do you get Opus level for the 256GB M5 Max?

Opus is probably ~2T parameter model, so that would probably not run on these. More like Sonnet.

Sonnet is estimated around 1T, so that is far beyond what's practical as well.

The 512GB could run GLM 5.3 which is Opus level

GLM 5.2 in NVFP4 is 465 GB. It would be a tough fit.

For local LLMs with a Mac, rule of thumb is you always want an Ultra (due to memory bandwidth). Even an M1 Ultra is superior to an M6 Pro in this regard.

There are no configurations even close to running something comparable to frontier model variants, they're simply far too large, but something like full precision Qwen 35b or DeepSeek 70b at 50+ t/s is well within available configuration, and potential for plenty of room for large context sizes.

256GB is enough for DSv4 Flash, expect maybe ~30tg/s, and a lot better profile.

I'm using Flash heavily, and I would describe it as nearly as intelligent as Sonnet-class in agentic coding, but more usable. Less world knowledge of course, and definitely a bit less intelligent; but not _that_ much.

On usability: Takes less handholding, less likely to make unsolicited refactors or whatever, and the writing style is readable.

It's not great at super-long-horizon goals as the Claude 5 models are; but if you have a good harness, you can get around that.

With the newest Qwen 3.8 27B model you can get opus on just about any new Mac.

It can fit into 16GB macbook ?

It can but not well, you really want at least 24 GB and preferably 64 GB for the full model

https://ofox.ai/blog/qwen-3-8-27b-run-locally-vram-gguf-2026...

I ended up getting an M5 Max MBP with 48GB and it runs Qwen3.8 in Q4. I developed 3 apps so far, medium complexity, with no issue using OpenCode. This is probably the minimum setup for comfortable local agentic coding in my book.

Anthropic can't have their IPO soon enough (rumored October) if these Macs match their reported specs.

Maybe $17k for a 512gig system that can do 1.2TB/s seems like a pretty good deal for a small office.

> M6 supports up to 32GB of unified memory to multitask across demanding apps and run LLMs on device for secure and private agentic tasks. It also provides up to 170GB/s of unified memory bandwidth — a 10 percent increase over M5 and a 2.5x increase over M1.

Isn't 170GB/s slow for bandwidth?

It is. That's the mac mini. For local LLMs you would want the Mac Studio, which tops out at 1.2TB/s.

Kinda. Strix Halo does 256GB/s of memory bandwidth, and is significantly slower than an M5 Max (614GB/s). Feels like intentional market segmentation?

You can get the m5 pro in the Mac mini with 307GB/s at 64gb of memory it’s $2899

It's fast by computer standards and excellent for entry level chip. The Pro/Max/Ultra chips are always faster.

Compared to something like VRAM it's slow.

It's actually quite fast, more bandwidth than NVIDIA RTX 4090 at 1.0 TB/s according to https://www.digitalocean.com/community/tutorials/gpu-memory-... Compute power will still be different, though.

The thread you're replying to is asking about the 170GB/sec memory bandwidth of the base M6, not the faster spec of the Ultra chips.

For max memory bandwidth you need to buy the Ultra versions (M5 Ultra: 1,2TB/s, this gets comparable to real GPUs regarding the memory bandwidth).

It's higher bandwidth than any dual-channel DDR5 desktop machine, but Apple never quote the memory latency, so hard to compare otherwise.

Would that be a difficult comparison to do fairly sine one is SoC and the other isn’t?

I meant that it's a difficult comparison between platforms without having all the details like memory latency, how the memory controller load increases latency (can increase by 4x on some platforms), cache latency, TLB entries, etc.

Extremely high bandwidth is great for copying data, but not as relevant for walking chains of pointers, where latency/cache/TLB entries are more important.

So just saying "high bandwidth == better" is true when other variables are the same, but they rarely are, especially in comparison to x86-64 offerings.

All of these are independent that it is a SoC with on-package DRAM.

You would want to get the M5 pro version with 307gb/s if you were interested in running local LLMs.

M6 and M5 Ultra both have comparable memory bandwidth per GPU core. I think they will perform well.

Here's me trying to justify this when I can run frontier models in the cloud for less than the monthly finance charge for this beast.

If you like to experiment with training / finetuning / etc on LLMs, these are actually incredibly ‘cheap’.

1.2TB/s memory bandwidth unlocks a lot with 256GB unified, and agentic AI is pretty good at optimising performance.

For comparison, to get 256GB with NVIDIA, you’re looking at a DIY workstation build (need pcie lanes), and like $70k?

The spark’s ~250gb/s bandwidth doesn’t really count here.

To experiment, its still al ot cheaper to prepare everything locally and then just rent a GPU Node on all of these non hyperscalers.

1-2$ / hour.

I'm not regretting my setup at home as it got paid by my company which makes sense here, but paying for electricity is quite high and makes already 0.3$/hour alone.

I would argue, the most interesting use case for running it at home is some personal agent which you want to run 24/7.

Neither is the right alternative to compare to. You aren’t going to hit 100% utilization (if you are, ignore me, this doesn’t some to you, and write a blogpost for me to read and share).

The comparison should be against renting in the cloud for the duration of your task for training and research or using pay-per-api-call providers for general inference instead of buying your own hardware (and paying the electricity and cooling bills on top), because let’s face it, the models you want to use are probably the same ones available on inference providers (but, yes, some are more trustworthy than others).

Speaking as someone that does ML/AI research, you are essentially paying a huge premium for being able to just run your Python script at any time without setting up a deployment script and harness to run the job remotely, while your hardware sits essentially idle the rest of the time.

The only way to make the math work is if you rent your hardware in the background for inference while you’re not using it in anger, but despite all the startups and promises that has never become as streamlined as mining bitcoins or shitcoins used to be and they don’t pay out as much as they say they would. Renting your hardware for training is another option but doing that is a lot more involved, options are fewer and farther in between, you won’t get as much utilization out of it, and doesn’t let you feasibly abort running tasks at a moment’s notice.

My card (RTX Pro 6000) is always doing something all the time from my queue; like some synthetic dataset generation up next. I still actively use runpods and openrouter for scaled stuff, I was spending a bit and then did the maths, and invested in it.

The maths to me was basically equivalent to prepaying for 242 days of runpod pricing for the same GPU; and I reckon I'd be able to get 6+ years of use out of this card with 96GB.

Plus there's the resell value -- it's actually appreciated by ~50% since I bought it.

Plus I do really enjoy that it's 100% local. I wouldn't feel comfortable giving my agents this much information if inference wasn't 100% local.

I wouldn't get another one, I wouldn't have as much value, but one is definitely paying off for me on the financial side.

Yeah not really lol.

Only reason to buy this if you want to own your compute.

Experimentation and inference are all going to be cheaper on the cloud

No, you need more compute for those use cases. That's why everyone trains on GPUs.

Why doesn't Apple make a tower PC with a standard ATX motherboard, power supply, extension cards, etc, like in the old days? That's immensely more upgradable, easier to cool, and I don't believe most people expect their desktops PCs to be tiny.

They sorta did right? It didn’t sell well at all.

I'm actually coming around to realizing that I _do_ expect a desktop PC to be small. I'm finding myself only considering Mini ITX desktop cases when I think about upgrading my desktop. I only have one video card and don't have storage requirements that necessitate lots of 3.5" drives, a big nvme is all I need.

We no longer need CD/DVD/floppy drives, storage has shrunk/moved to the cloud. The only thing that's really grown inside a pc case is the video card, and most of these ITX cases are built specifically around fitting popular cards.

Even folks primarily focused on gaming are probably thinking that a full ATX case is a lot of wasted space.

Maybe it's just me, but I think ATX full and mid towers are going the way of the dinosaur.

Personally I'd love to have a smaller PC for gaming, but I've been scared off by Mini-ITX supposedly being harder to work with and tending to have worse cooling and therefore worse noise. I wish PCs were less fiddly to put together.

Mini-ITX is harder to work with and usually does have worse cooling.

There's just less space, so it's harder to get things in and out. And you have a smaller cooler because there's less space, at least for air cooling.

I haven't used itx with video cards, but that's hugely constrained...

I have a 9950X3D and an RTX 4090 in a 10L case (FormD T1).

Water cooled but it does get a bit loud when I play games, silent for regular usage.

But then how will you fit your action figures and extra displays that go in your case?

They were obsolete a decade ago, but priced out by artificially expensive "SFF PC" cases, fans, and power supplies. PC building completely transitioned to fashionable cargo cult in the late 2010's, with the rise of twitch.

> Maybe it's just me, but I think ATX full and mid towers are going the way of the dinosaur.

Even standard ATX cases are kind of small if you want to build a full custom loop and need space for radiators, pumps, reservoirs, tubes, etc. My O11 XL is admittedly enormous, but it is under my desk and does not occupy any space that wouldn't otherwise remain unused.

As for storage, I categorically refuse to store things in the cloud. My files belong at home on my hardware, not on someone else's computer in the "cloud."

The reason the mini & studio work as well as they do for the use cases that have made them popular again is because the architecture puts memory, cpu, and gpu/npu on the same die, and puts storage right next to it, which is why you get the storage & memory bandwidth you do. The separability of those components would break that, which is why the Mac Pro didn’t survive past one generation in the Apple Silicon era.

The ram is not on the same "die" it's in the same package

Distinction without a difference in this case...

SSD controller is on the same die, but the actual SSDs in a Mac Studio are slotted and technically upgradeable. Storage bandwidth generally matches NVMe drives of the same generation, as Apple is ultimately using the same NAND modules over PCIe.

My ideal Mac Studio would have an NVMe slot or two for additional internal storage, and those slots could be accessed without opening the case. But Apple won't do that, as convenient internal storage upgrades would likely lower their profits.

The memory and GPU are integrated into the CPU so those can't be upgraded anyways. That's also how the memory can be so fast (shorter physical distance).

Because it doesn't look cool?

Why would they / what would the consumer value be?

You can't upgrade the processor because it's soldered. You can't upgrade the ram because it's soldered. You can't upgrade the SSD because it's soldered / bonded to the CPU. You could maybe have some PCIe slots, but not many drivers for macOS, so what's the point?

Yes, it was different in the old days, but Apple is ever more a closed hardware architecture with few options. If you want choices, Apple is not for you.

I'd love for Apple to bring back the Mac Pro, but Apple hasn't had an affordable Mac Pro tower since the 2012 model. Apple got rid of the tower in 2013 in favor of the "trash can" model. When the trash can languished for years without updates due to limitations with its design, Apple brought back the tower Mac Pro in 2019, but at a dramatically higher entry MSRP, which priced out many users of tower and trash can Mac Pros (I have a "trash can" and was very disappointed by the 2019 Mac Pro's entry MSRP). The introduction of the Mac Studio revealed a large price gap between similarly-specified Mac Studio and Mac Pro models. Given the lack of expandable RAM in Apple Silicon Mac Pro models, and also given the lack of support for NVIDIA GPUs in macOS, there weren't a lot of customers for Apple Silicon Mac Pros when the Mac Studio was available at a lower price, and thus the Mac Pro got discontinued.

In many ways, the Mac Studio is the spiritual successor to the trash can Mac Pro. Apple Silicon's architecture solved the thermal problems that limited the trash can Mac Pro.

They used to support that. Isn’t possible to do unified memory in those configurations? If not, then I expect it’s a headache Apple does not want to deal with. For example, earlier versions of Metal have unified/discrete memory modes with different API availability and behaviour in each - what a nightmare to support and maintain and keep bugs out it would be

> That's immensely more upgradable

This is the exactly reason.

For this reason: more upgradable

Apple wants to sell you a new system every couple of years, they don't want you to upgrade. They want you to throw it away the system completely and replace it (and they claim they are environmentally sustainable, btw) when in the PC world you would replace RAM, CPU, GPU, motherboard, maybe the SSD, but surely you can keep the power supply for decades, keep cooling fans, not talking about the chassis that could last 50 years because it's a just metal box.

They are a rip off these products, and I don't understand people that buys them.

BTW even when Apple did produce towers like Mac Pro they did so in a purposely non-standard way, such that you couldn't, for example, replace the PSU with a standard ATX one, the only standard components were really RAM, HDD, CD-ROM drive and in Intel models the CPU, and of course PCI expansion boards (sort of, because to work GPUs needed to have an Apple specific firmware in the old PowerPC models). You couldn't for example upgrade the motherboard, if you wanted to change the CPU to a new generation you also needed to buy a entirely new Mac with a new chassis, new PSU, etc even if what you had were perfectly fine.

Bizarre there isn’t a 1TB RAM option hidden away for the excessively frivolous or VC funded.

Same reason they cut the big options on the existing models, this way they can sell more devices. The additional cost for the additional 512GB would have to make up for the loss of another sold device otherwise. No idea if there would really be that many people buying this then while on the other hand AI stuff makes people do crazy stuff, so...yeah :)

Considering Apple pricing it actually might.

256GB model is $10k and the 512GB version will probably be double

They seem to be suffering from the supply constraints like everyone else. They phased out the higher capacities on the M3 Ultra Mac Studio a while ago, and if you order a 128GB MBP, say, you're looking at six weeks or more for delivery.

It's worse than six weeks: Apple quoted me 10-12 weeks yesterday for an M4 Max Studio (about Nov 17) December for the M3 Ultra. Planning to lock in a new model soon. I went looking for used but those are wildly expensive. Strange times. https://hard.bargains/posts/mac-before-sept-22/

It's more bizarre that Apple got caught with pants down. Focused on CPUs and ignored RAM.

Seems like miscalculation. If they had their own fab for RAM, they could completely corner the market today.

They have no fab for CPUs, they are manufactured by Samsung and TSMC. The bottleneck is in manufacturing RAM not CPUs so there is nothing Apple can do here.

Let me rephrase. They have not booked enough capacity at contract manufacturer.

Recently Micron pointed that finger and threw that shade in Apple’s direction saying that their priority for years has been squeezing margins in their favour, so there wasn’t the slack in the system to build the fab capacity that is now desperately required given the sudden need for RAM now that there’s a cause to use it in AI Inference.

They don’t have their own fab for CPUs either.

They would have had to start building that fab 5-10 years ago. It would have been incredible foresight to do so and it would have looked insane.

Boy, oh boy Apple is the new shovel seller during AI gold rush.

Most people at Apple have already realized that their processors are already too powerful for regular users - heck, as a developer my M2 Pro with 32 GB RAM is more than enough for me.

Regular users don’t care about local AI either. So, they will probably extract as much money as possible during AI gold rush, but then we will most likely see Apple

a. Making their software worse (god forbid, forced updates)

b. Making their hardware impossible to repair (as they almost accomplished this already) and easier to break.

I dunno, Apple's been throwing many bones to their customers who use their Macs for AI stuff; like Apple working with, and signing TinyGPU's NVIDIA eGPU drivers.

Plus introducing features like RDMA over thunderbolt, which is critical for distributed inference/training/etc. On the software side, Apple is investing heaps.

It's still ridiculous they don't support expandable NVMes, but the memory being soldered makes sense, you need it for 1.2TB/s bandwidth.

They are still selling high-margin hardware. Apple loves selling high-margin hardware.

Wow, didn't know this exists. So instead of ditching my M1 Pro Macbook for a newer one with more RAM and power I could do this instead (if reasonable regarding pricing).

EDIT: AMD too, it's not limited to Nvidia, nice.

> Boy, oh boy Apple is the new shovel seller during AI gold rush.

They're eating nvidia's lunch.

Incredible that LMStudio, an open and ever evolving (unstable) project, is cited as a speed reference.

It's not open source.

https://en.wikipedia.org/wiki/LM_Studio

It’s license may be proprietary but delivery is fairly open given https://github.com/lmstudio-ai

Mac Studio with M5 Ultra starts at $5,499 - computing this expensive belongs in a server room serving multiple people not on a single person’s desk. (But still cool).

M6 (base) in Mac Mini too:

https://www.apple.com/mac-mini/

So weird, launching M6 so silently.

Allegedly, they aren't making a big deal out of the M6 which won't have Pro, Max or Ultra configs at all, instead waiting for the M7: https://9to5mac.com/2026/08/08/apple-m7-chip-heres-why-it-ma...

Interesting, thanks for sharing.

512GB unified memory option coming available in October.

The 256GB option is +$4,000 - the overall price for 512GB setup would probably be $20k!

It needs to be a little less than double the overall price for 256GB option, otherwise it's better to get 2 256GB and link them.

you can't link onchip memory.

True, but you can link them up over thunderbolt or Ethernet. If your goal is to run local LLMs, not all weights need to live on the same computer. You can segment the workload by layers and pass the activations along the lower bandwidth interconnect with a small perf penalty. Also you get double the CPU/GPU cores allowing for better multi user/agent performance.

Ah sorry you are right I missread what you meant by it.

I was thinking the same as you as far as price per value, it does make seance to get 2x of these things IMO, but what throughput hit would you see in linking versus one machine? latency does matter, and there must be a trade off no?

The new mac minis and mac studios are going to be in shortage for at least first 6 months from 9.22

It was not long ago all the illustrations in an M chip press release were graphs showing it flouncing the competition. Now it's just random screenshots of desktops. Skimming this I didn't see a single claim of performance advantages over competitors. The massive thing Apple still needs to solve is making sure customers don't miss being able to plug the latest greatest GPU into their tower, they should not be resting on their laurels.

Waited years for a decent Desktop lineup like this as I was hugely disappointed in the generation old M3 Ultra when it was released. Unfortunately it took Apple so long that the memory prices are now through the roof - USD $4000 for 164GB RAM upgrade is insane.

I may consider a M6 Mac Mini as a stop-gap whilst waiting out RAM Apocalypse to be over. Basically abandoning any ambitions of AI sovereignty and riding out subsidised LLM pricing for the next couple of years.

Assuming future models will be any cheaper is not a good gamble lol

I was looking forward and hoping that the Mini and Studio would have 8K at 120Hz. Oh well, maybe the M7s will have that.

It seems super reasonably priced to me. It's only twice as expensive as my first Mac which only had 128K of memory.

That came with a monitor and floppy drive.

and this one comes without a floppy drive.

Base models are okay-ish.

But +4000$ for an additional 128GB of ram is simply milking the customers, as they know they will have many of them.

Still absurd, but it's $4K for an additional 164GB RAM.

Pretty much every company milks their wealthier customers and it frees them up to be more competitive at the lower end. Also, if they charged a more reasonable amount, they would get backordered very quickly.

Depends on your perspective. People think nothing of spending 50-100K on a car that basically gets them to work. But the thing they use for day to day work then gets the evil eye when it costs more than 1K. It's slightly irrational. Not everybody needs a high end mac. But when you do, it sure is nice that you can get one.

I don't actually own a car and my startup is bootstrapped and our salaries are modest. But the one thing we spend on is laptops. I have M4 max pro with 48GB. That thing was on the expensive side (~4.5Kish). But it delivers a lot of value and I spend most hours I'm awake using it. I like fast builds. I like that I can try out open source AI models. And I like just having the option to run those.

We actually lease them and mine costs something like 105 euro/month. Including Apple Care. I don't need a Mac Studio but I could see some roles where that would not be a crazy expense. Even the tricked out version that basically only costs the same as a very modest car.

Seems they’re actually advertising pcie expansion. Dang I’d love to switch to a Mac for my cuda kernel writing.

Not a surprise this is launching now. John Ternus is taking the reigns as CEO in a week and was previously SVP of hardware engineering.

I can't hate a direction where Apple becomes more about building great computers rather than trying to force more and more subscriptions. I do wish they would fix many of the long-standing OS and native app problems.

If I were to get one of these, realistically what is the most advanced AI model I could run locally?

GLM 5.3 in ~q4 quantiziation with 4bits

Insta-leased. I’ve ran over 1 million prompts on ollama to determine “is X website”. I’m ready to use use bigger models locally to do more thinking now. I have zero desire to give Dario anymore than $20 a month after his “you all are gonna be mass unemployed” spiel.

How does this compare to NVIDIA DGX Spark? From what I see, if the performance is better than 2x DGX Spark, this could boost and in a way reshape Local AI in a significant way (along with all the OSS models wave that's happening) - look forward to that!

Their statement about sharing an Ultra for an office LLM server is a little misleading. The tokens/second would be too slow for any business use case. You'd want to buy multiple GPUs for that purpose, probably Nvidia, despite the extra cost.

Nice for them to add Thread and 10G ethernet to base Mac Studio. If they allowed first class Linux, this could be a great home server machine. I ordered the base model. 512GB SSD scares me but would I even notice plugging in TB5 external drive?

Shocking pricing.

Ugh. I was waiting it out on a very old Mac, but the pricing these days is insane, so now to wait it out for at least two years more and hope for better pricing or just suck it up and pay a lot for less.

512GB unified RAM is going to be really good for running local LLMs.

What is unified ram? RAM and VRAM as one thing?

Yep, exactly. It's just one shared pool of RAM with zero-copies necessary.

Big ole pool of very fast ram that can be accessed by the CPU and GPU. Lets you run larger models. AMD does the same thing with Strix Halo. I have a 128gb machine at home, and have had difficulties running 120b models, but 70b and below run pretty well.

AMD Strix Halo unified memory still requires allocating X to GPU, x to CPU at UEFI.

So not as flexible as apple's unified memory.

Not in Linux! See setting GTT: https://rocm.docs.amd.com/en/docs-7.2.0/how-to/system-optimi...

> Big ole pool of very fast ram

It’s more that it’s a very parallel architecture than fast.

The LPDDR5X under the hood is slower than what you’d find in a graphics card vram of years ago

That’s why you get consumer macs with 512gb while GPU makers are reluctant to give you more than 16gb unless you pay dearly. It’s not the same kind of mem

In best case its also on-chip high speed ram.

Yes

It's game changer but it's too bad without enough memory

Eff AI! My next computer purchase will probably be in a decade!

8000 dollars for a 16TB harddisk? At that price point I can load out two full PC’s for the price of a single disk, even at these inflated prices.

Just insane RAM price. Can’t wait for the bubble to pop. AI is useful but not at the currently unsustainable price (too expensive for the provided value to corporate customers at too cheap per token vs the actual cost of creating it)

You’ll be waiting a while. Ram capacity is bought out for 2027

Anyone considering the lease option? Anyone done it with Apple before? Could it be tax advantaged?

Unless profits no longer matter this is the peak of computer pricing. It will go down from here

unfortunately it does not run linux

Containers and VMs are well supported.

I’m running an Ubuntu VM via Lima on my m4 Mac mini. Works great.

It’d be great when they start releasing these machines with the local models ready for use!

They already do: https://developer.apple.com/documentation/foundationmodels/s...

Also see: https://apfel.franzai.com

buy a nice BMW 3 series or buy a 512GB m5 ultra... ;-)

I'm glad they brought the 512GB option back.

Really want to get my hands on a m6 32GB 2TB model but don't really have a need for one.

No looking to cheat, which is better: the MAX or ULTRA?

Depends if it's Pro Ultra Max, or Plus Max Ultra.

it cost a fortune with current memory cost. but i was waiting for this! whats the bandwith compared to a rtx pro 5000?

M5 Ultra in a Mac Studio form factor is exactly what the creative pro crowd has been waiting for.

The unified memory architecture keeps paying off. More bandwidth means bigger local models.

Am I the only one that now finds press releases like this similar to "AI Slop"

I know there's tons of marketing language, buzz words and attempts at convincing me of some agenda that isn't super clear without lots of effort in "validating" the slop. I guess its not bad "slop" though if a human put in effort in editing it (imo >50% human curating = not really bad ai slop)

Though I still would prefer I could just get the prompt. What human thoughts, direction and "prompt" went into writing this article? in the same way as we ask for the prompt for AI generated outputs, I would prefer it for human generated output too. For writing at the least. I could have saved time, got the purity of the argument, and got more clear information. I wonder if we can get a future where humans just express their intent with each other and stop trying to hide our agenda; I want a world we can trust each other greater and interpret and act on our goals without the noise of trying to impress or market to each other & the additional words that go into that.

"512GB memory option for M5 Ultra coming late October"

anyone know if this is "pre-order" is coming late October, or "will be available to ship in late October, thus pre-order will be available earlier than that"

Why do I only have two kidneys dude

Very well timed for John Ternus's first quarter.

Very much happy with my machine.Got a base m4 max studio in December. 2350eur what a deal

I only with I got more than 96GB at the time.

$5500 for 96GB of RAM? insanely expensive (the MacOs is really bad compared to Windows or Linux).

The article's headline contains an em dash. I wonder if it's AI-generated

To me it seems like a cheeky way to preface the AI part of the title.

My sarcasm detector just made a small cloud of smoke. What does that mean, a buffer over or underflow??

Reply on news.ycombinator.com