MacBook Pro with M5 Pro and M5 Max

Nothing has broken and I consistently get 4-6 hours of heavy work time while on battery. An amazing machine for the price I paid.

As there target for that marketing, I can report it hits home!

But objectively, there is nothing wrong with my current experience at all.

I have never had that experience over many generations and types of machines. The M1 keeps looking better and better in hindsight.

—-

Looking forward, either the M5 is the next M1, a bump of good that will last. Or Apple will be really firing on all cylinders if it can “obsolete” the M5 anytime soon.

I thought the buyer was insane to buy it at that price. But, of course mine had a decent spec and still had the Apple care warranty with very low battery cycle count. After the sale, the buyer told me the truth: The M1 is the best chip Apple ever made and I wouldn't see much of a difference in real world between the M1 Pro and an M3 Pro unless it was the Max version of the chip.

I didn't believe him then. But, after a year of being on M3 Pro, I gotta say he was spot on. Don't get me wrong, the M3 Pro is definitely faster in a lot of things. But not 3x or 2x faster like Apple always like to market. I can open a few extra tabs without slowing down, compile times (Elixir) did get somewhat faster. But definitely not faster to the point where there were two generations worth of performance improvements like Apple claimed.

The M1 chip series is vastly underrated.

Also, my wife's still using the older touch bar MBP, and we'll, it works fine for her too.

I'm not sure who needs the newer pros.

>I think I will get another 2-5 years out my mine.

I only own a M4 because the M1 had a hardware fault and I needed a replacement ASAP. (I sold the M1 after repair.)

Although I'm glad to have a newer machine with longer future support, I have yet to notice any meaningful performance difference.

the air series is really good, and very light

my M1 is now noticeably heavy and I don't think upgrading to another Macbook Pro is the move the resell value of the M1 did not hold, specifically the bumped up storage models. There doesn't seem to be a market for 8TB of space specifically, but the base 1 - 2TB holds its value because the baseline of the MBP holds its value

M5 Max looks tempting if there is a very compelling tradein, but the M1 Max is pretty old so I don't have real hope of that, but I'll look. For AI Inference the difference doesn't seem good enough yet and necessary enough. I'll still need to use the cloud or aspire to have a specialized machine with more RAM or circuitry on my network.

Of all the stupid things Apple has said lately, this is the most obtuse, pro-market-insulting nonsense. Intel-based Macs were knee-deep in issues Apple wasn’t fixing, and then along came the snappy and always-cool M1.

That was the best time for customers to upgrade. The new Silicon generations can be quite good, but they’re not worlds ahead in anything.

I’ll upgrade my M1 when Apple releases a macOS worthy of being used by its pro customers.

What is tricky is not even CPU/GPU, but that in a Macbook it is impossible to upgrade RAM (easier to understand, as it is tied to the processor), but also the hard drive. Correct me if I am wrong, but I bet it is a decision by Apple, so people buy newer Macs more often.

This does look like a nice machine though.

I’ll probably wait for the M6 Max. If/when RAM comes down they might stuff 192 or 256 gigs in one, which would make it able to run larger tier open weights models.

128 is kind of an uncanny valley for models. Bigger than you need for the mid tier and too small for the huge ones.

I keep telling people that the best laptop value on the market right now is to buy a refurbished MacBook Pro M1/M2. I stand by that from a usability and performance standpoint, but I feel weird about recommending a laptop that could only get security updates for another 3 years.

My work laptop is an M1 Pro and it is also doing totally fine. At work we used to do laptop upgrades on a 3 year cadence but the M-series laptops are so good that we switched to 5 years instead.

They probably can't do that because of potential patent issues that might surface.

Lawyers say no.

Oh dear 14B and 4-bit quant? There are going to be a lot of embarrassed programmers who need to explain to their engineering managers why their Macbook can't reasonably run LLMs like they said it could. (This already happened at my fortune 20 company lol)

https://www.apple.com/macbook-pro/

0: https://entrpi.github.io/eemicrogpt/

Batch-1 token generation, that is often quoted, does not benefit from this. It's purely RAM bandwidth-limited.

M5 128GB RAM with 614GB/s memory transfer

This is a huge step over M4 32GB 153GB/s memory transfer

For local LLM this make it a replacement for a DGX Spark, which offers a third of the transfer speed and is not something you toss in your backpack as your laptop. It’s practically useful for a lot of local use cases and that I think is the 4x factor (memory xfer) - but the 128Gb unified headroom tremendously improves the models you can run and training you can do.

I don't mind it, I open Apple stock. But I'm def not buying into their rebranding of integrated GPU under the guise of Unified Memory.

Now extrapolating in line with how Sun servers around year 2000 cost a fortune and can be emulated by a 5$ VPS today, Apple is seeing that they can maybe grab the local LLM workloads if they act now with their integrated chip development.

But to grab that, they need developers to rely less on CUDA via Python or have other proper hardware support for those environments, and that won't happen without the hardware being there first and the machines being able to be built with enough memory (refreshing to see Apple support 128gb even if it'll probably bleed you dry).

Neural Accelerators (aka NAX) accelerates matmults with tile sizes >= 32. From a very high level perspective, LLM inference has two phases: (chunked) prefill and decode. The former is matmults (GEMM) and the latter is matrix vector mults (GEMV). Neural Accelerators make the former (prefill) faster and have no impact on the latter.

I assume they have a moderate bet on on-device SLMs in addition to other ML models, but not much planned for LLMs, which at that scale, might be good as generalists but very poor at guaranteeing success for each specific minute tasks you want done.

In short: 8gb to store tens of very small and fast purpose-specific models is much better than a single 8gb LLM trying to do everything.

Apple is in the hardware business.

They want you to buy their hardware.

People using Cloud for compute is essentially competitive to their core business.

Remains to be seen how capable it actually is. But they're certainly trying to sell the privacy aspect.

So as most people in or adjacent to the AI space know, NVidia gatekeeps their best GPUs with the most memory by making them eye-wateringly expensive. It's a form of market segmentation. So consumer GPUs top out at 16GB (5090 currently) while the best AI GPUs (H200?) is 141GB (I just had to search)? I think the previou sgen was 80GB.

But these GPUs are north of $30k.

Now the Mac Studio tops out currently at 512GB os SHARED memory. That means you can potentially run a much larger model locally without distributing it across machines. Currently that retails at $9500 but that's relatively cheap, in comparison.

But, as it stands now, the best Apple chips have significantly lower memory bandwidth than NVidia GPUs and that really impacts tokens/second.

So I've been waiting to see if Apple will realize this and address it in the next generation of Mac Studios (and, to a lesser extend, Macbook Pros). The H200 seems to be 4.8TB/s. IIRC the 5090 is ~1.8TB/s. The best Apple is (IIRC) 819GB/s on the M3 Ultra.

Apple could really make a dent in NVidia's monopoly here if they address some of these technical limitations.

So I just checked the memory bandwidth of these new chips and it seems like the M5 is 153GB/s, M5 Pro is ~300 and M5 Max is ~600. I was hoping for higher. This isn't a big jump from the M4 generation. I suspect the new Studios will probably barely break 1TB/s. I had been hoping for higher.

• Studios with Ultra Mx, now 4-way RDMA over Thunderbolt 5, and enormous RAM and SSD options, suggest a strong focus. I don't know what else that RAM would be intended for. Four Studio Ultras (total of 360 GPU cores with M5 Ultras?) with 2TB of unified RAM is a local model beast.

• They refashioned their GPU cores to better support both graphic and neural processing, despite already having focused NPU cores.

I would say they have been leaning into local models for several years.

I expect we will see more models being optimized for smaller sizes, as demand for them increases. With hardware performance and neural focus trending up, and model requirements/quality trending down, the next few years will be interesting times.

What would make me happy: Ultra x 2 (i.e. 2xUltra, 4xMax, 8xPro, 16xM5) packaging in the Studio. With 8-way RDMA. Mac Kong. Perhaps Apple will start making server cards again.

  Are they doubling down on local LLMs then?

Apple's stated AI strategy is local where it can and cloud where it needs. So "doubling down"? Probably not. But it fits in their strategy.

I think I'll pass on upgrading.

Honestly, I think that's the move for apple. They do not seem to have any interest in creating a frontier lab/model -- why would they give the capex and how far behind they are.

But open source models (Kimi, Deepseek, Qwen) are getting better and better, and apple makes excellent hardware for local LLMs. How appealing would it be to have your own LLM that knows all your secrets and doesnt serve you ads/slop, versus OpenAI and SCam Altman having all your secrets? I would seriously consider it even if the performance was not quite there. And no need for subscription + cli tool.

I think apple is in the best position to have native AI, versus the competition which end up being edge nodes for the big 4 frontier labs.

Do think it'll be common to see pros purchasing expensive PCs approaching £25k or more if they could run SoTA multi-modal LLMs faster & locally.

"AI" (LLMs) may or may not have a bubble-pop moment, but until it does Apple get to ride it on these press releases and claims. But if the big-pop occurs, then Apple winds up with really fantastic hardware that just happens to be good at AI workloads (as well as general computing).

For example, image classification (e.g. face recognition/photo tagging), ASR+vocoders, image enhancement, OCR, et al, were popular before the current boom, and will likely remain popular after. Even if LLM usage dries up/falls out of vogue, this hardware still offers a significant user benefit.

I just don't get why they're dropping the ball so much on this.

I love the push to local llms. But it’s hilarious how apple a few years ago was so reluctant to even mention “AI” in its keynotes and fast forward a couple years they’ve fully embraced it. I mean I like that they embraced it rather than be “different” (stubborn) and stay behind the tech industry. It’s the smart choice. I just think it’s funny.

So yes, the LLM should be inferencing on your prompt, but it should also be inferencing on 25,000 other things … in parallel.

Those are the compute needs.

We just need compute everywhere as fast as possible.

This correlation of Apple and privacy needs to rest. They have consistently proven to be otherwise - despite heavily marketing themselves as "privacy-first"

https://www.theguardian.com/technology/2019/jul/26/apple-con...

No change from the previous models then, 16GB->32GB was already $400. They're cutting into their previously enormous margins to keep the prices stable, rather than hiking the prices to maintain their margins.

Isn't this it?

My M3 Pro from a few years ago for the same price had 18GB.

Interesting that this hasn't budged since the memory shortages appeared.

Unless you are planning to do some serious inferencing, or complex multi-agent setup, then you don't need the memory.

this is my exact opposite experience. my M3 Max from 2 years ago now has <2hrs battery life at best. wondering if any experts here can help me figure out what is going on? what should i be expecting?

Incidentally, I just switched to Asahi Linux, but that was for software quality and openness reasons, rather than anything to do with performance.

Unfortunately, number always must go up (and the rate at which the number goes up, also must go up).

But I think this predates Tahoe.

The new tensor cores, sorry, "Neural Accelerator" only really help with prompt preprocessing aka prefill, and not with token generation. Token generation is memory bound.

Hopefully the Ultra version (if it exists) has a bigger jump in memory bandwidth and maximum RAM.

Wondering if local LLM (for coding) is a realistic option, otherwise I wouldn't have to max out the RAM.

This seems even likely as the memory bandwidth hasn't increased enough for those kinds of speedups, and I guess prefill is more likely to be compute-bound (vs mem bw bound).

Also can you run batchwise effectively like vllm on cuda?

Enough to run multiple agents at the same time with throughput?

I think this is a new design, with Apple having three tiers of cores now, similar to what Qualcomm has been doing for a while.

I think how it breaks down is:

- "Super" are the old "P" cores, and the top tier cores now

- "Performance" cores are a new tier and seen for the first time here, slotting between "old" P and E in performance

- "Efficiency" / "E" are still going to be around; but maybe not in desktop/Pro/Max anymore.

The base M5 has super/efficiency cores.

The Pro and Max have super/performance cores.

  Whoah, both the Pro and Max CPUs feature 18 cores. This hasn't happened since M1 Pro/Max. This is a surprise.

I believe they lower the clock speed, limit how much work is done in parallel on each core, and limit how aggressive the speculative execution is so less work is wasted.

128 GB maximum.

Sigh.

It's the first time I've ever been so repulsed by a design that I actively avoid it just... out of sheer preference.

If you move your home directory to a different disk partition, you can even share it between two different macOS versions!

The high memory Macs have been great for being able to run LLMs, but the prompt processing has always been on the slow side. The new AI acceleration in these should help with that.

There are also workloads like compiling code where I’ll take all the extra speed I can get. Every little bit of reduced cycle time helps me finish earlier in the day.

And then there’s gaming. I don’t game much, but the M1 and M2 era Apple Silicon feels sluggish relative to what I have on the nVidia side.

   and that’s with doing pretty demanding 3D and LLM work.

Before:

"We have 6 performance cores and 12 efficiency cores"

After:

"We have 6 super cores and 12 performance cores"

"Wow, how did you achieve this?"

"We changed the names."

https://appleinsider.com/articles/25/10/15/eu-gets-what-it-a...

In the US they provide one in the box free of charge.

So, if you want one of mine, you can have one. On me. Because I'm fucking drowning in the things and appreciate not having to deal with another one.

The EU requires that users must be able to buy a device without a charger. It's a huge supply chain challenge to add two variants of every single SKU, one with a charger and one without. So the obvious solution is to sell the charger separately, since you need that regardless, and always sell the device without a charger. You avoid having two variants of everything that way.

Now, you could maybe argue that Apple should default to bundle a charger with your laptop, so that you'd have to uncheck a "bundle charger" checkbox on their website. But do you really care whether your laptop costs $2200 and you can buy a charger for $60 or your laptop costs $2260 and you can save $60 by removing the charger?

You can make an argument that doing it Apple's way hides a price increase. And yeah, that's probably fair. But it's not like Apple is afraid of non-hidden price increases either.

The only differences that are more expensive EU vs US is the AppleCare+ and taxes.

US looks like you pay yearly for AppleCare+ while EU it has to be for a fixed number of years.

I think I read somewhere long time ago that Capture One is also using Qt for GUI, though cannot find this anymore, so probably not true.

You sadly just missed the window or cancelled too soon.

Normally if your current order is in progress they swap it out for the best closest spec for the exact same price you ordered the M4.

The prompt processing sped up.

Not the output generation.

M4 was notoriously slow at this compared to DGX etc.

Now it starts at $1699, a $100 bump but comes with a 1TB SSD. Previously it would have cost $1799 for the 1T SSD, so it's a $100 bump on base price but you are also getting 1TB SDD for $100 less than before.

Oh really, it's universally better?

> For those of us with astigmatism it's really night and day experience.

Oh. So it's better for someone else with a specific eye condition, who is practically guaranteed to never use a MacBook that I buy?

It's chiplets just like GB10, Strix Halo, etc. One die has the CPU and the other die has the GPU.

How is that different from the silicon [bridge] they were using before?

It's probably similar.

the two dies don’t have to fabbed next to each other

They never were; this is a widespread misunderstanding.

But would this affect the interconnect speed somehow?

Apple never documented the internal interconnect for the M4 Pro/Max and now they don't document it for the M5 Pro/Max so we don't know. It's probably better to read reviews and avoid theorycrafting and backseat driving.

They seem to market it as a technological advancement, which it is, but rather than being excited im actually worried about hidden latencies that could come with that approach. Have you found any interesting info on that yet?

Yes, different computers have different keyboards. Macs have had Mac conventions since 1984, and if you're not used to it you're not used to it. Instead of sticking to "the thing I'm not used to is wrong", I would suggest trying to be objective about which is the better UX.

I spent the first 25 years of my computer use being a Linux and Windows user and barely ever touching a Mac, so I've had to adjust, but to be honest, the truth is that Apple was always right and Microsoft made the wrong call.

Ctrl is meant to send control characters. This has been well defined since the birth of ASCII. In macOS, cmd-C is always copy, cmd-V is always paste. It does not matter what mode or program you are in.

Windows was designed for an IBM PC where the keyboard only had Ctrl and Alt, and when they copied Apple conventions (like cmd-X, C, V for cut, copy, paste) they made the wrong decision in using Ctrl for it. We've paid for this debt ever since. GNOME, KDE, XFCE, [...] devs continued this travesty by copying Windows, and so now on Linux Ctrl-C is copy in GUI apps, unless it's a terminal, in which case Ctrl-C will break out of your process, and copy is Ctrl-Shift-C. This is insane and bad UX.

The correct choice for a Linux DE would have been to use the "super" key ("Windows logo") for copy/paste etc, but of course they couldn't do that because 1) not everyone had that key and 2) it would confuse Windows users.

- Select with touchpad?? You must physically press its button like WTF

System Settings -> Touchpad -> Tap to click

- Mouse with backward/forward button?? Good luck!!

This is stupid. I will agree that Apple's insistence to not really support anything other than their own severely limited input devices is boneheaded. I recommend installing SensibleSideButtons.

Phones have less configurability, they sell more, and colors seem more important.