Hacker News Reader

Comment by jauntywundrkind | Hacker News Reader

Ampere is still the only people making a decent purchaseable ARM server. Not loving the idea of there being a bunch of new debt hovering over them now.

View on HN

Got a chance to play about with the Altra Max to do some batch jobs last year.

128 cores going 100% in htop was fun!

The main problem with those is the memory controller starts to degrade a lot when you use all 128 of those cores.

We were doing some testing and > 96 cores at 100% caused a massive degradation in performance. We ended up going with dual 32C/64T Epycs (which cost twice as much) as a result. If they fix it in the Altra One chips they will be back on the table though because they were very good power wise for our workload and quite price competitive in a supermicro chassis.

You can buy Nvidia Grace CPU only systems. Neoverse V2 is actually pretty solid.

https://www.supermicro.com/datasheet/datasheet-NVIDIA-MGX-1U...

Nvidia is targeting AI specific or fastest single core performance . While Ampere are targeting general Web CPU usage which is larger number of core per socket. The two are very different. I would have like Nvidia to enter this space as well but they seems to be focusing on everything AI related at the moment.

Or there are some specific workload on web that their CPU benefits and get filtered across ( DB hosting for example ) rather than going into a separate market for general usage.

What about

https://www.apple.com/shop/buy-mac/mac-pro/rack#

That's not a server. It is a rackable desktop.

You won't run general purpose server apps on this.

Not with that attitude you won't. Clearly the machine is the wrong color for your workloads.

What is a general purpose server app?

Does Apple prevent their users running applications which listen to outside connections?

I think the issue is that you'll have the base macOS system with GUI and all taking up resources.

Compare that to a bare metal install of Proxmox or similar.

Yeah, I know, but given that you leave the user not-logged in and set the items system-wide startup items, the memory pressure on a 64GB system can be seen as negligible with some mental flexibility.

I think the primary use for a rackable server like this is automating building and testing of macOS native, or macOS specific applications you develop, or build them. Or as a shared remote desktop host for specific tasks.

Also we had (recently deprecated) macOS Server package for quite some time. So it's not unsuitable for servers. It's more suitable for specific jobs as a server.

Installing Linux servers are too easy now, so I don't think macOS should evolve a "server-specific" version either.

Proxmox runs fine on M-series Macs. macOS also runs fine with no GUI.

Either way, it wouldn't be a good fit when Ampere exists.

For low power modular compute, the way GitHub and Amazon re-package Mac Minis is not the moest efficient way to get Apple hardware in a rack with remote management (and still no ECC), but it does have a performance per watt and compatibility for macOS CI workers that you can't get any other way.

Ideally, this stops being an issue when the ARM64 ecosystem can build those apps natively. We're pretty close on that one, and it means you can use Ampere and Graviton for that.

have fun when that cheese grater hangs in the datacenter and you have to physically go in there to push a button, connect a display, a keyboard and a mouse to do troubleshooting, because apple in their infinite wisdom has decided that a rackable mac should not have ipmi.

Why not use JetKVM [0], and a remotely switchable datacenter PDU?

[0]: https://jetkvm.com/

you certainly could... but when you're spending mac-pro money you might as well get a machine with built-in ipmi.

it's not that you can't -- it's just inefficient: you're spending more money for less features.

No BMC, severally limiting remote management, so operations teams responsible for physical servers don't want to touch it.

JetKVM is a remote KVM, not remote management. It does not have the same integration into the internals of a server.

But it comes with a ... checks notes ... magic mouse!

Upvoted to save you, out of pity for this terrible joke. :)

Great, I can still make the HMS Courage[0]

[0]https://www.igen.fr/accessoires/2015/11/non-la-magic-mouse-n...

They also won't provide any guarantees about components continuing to work, such as booting over thunderbolt that some clouds rely on. There's an awful lot of financial risk in providing MacOS as a service.

Apple has a BMC on the Mac mini, Mac Studio and Mac Pro. It's called Lights Out Management (LOM). https://support.apple.com/guide/deployment/lights-out-manage...

AFAICT LOM only has three commands "PowerON", "PowerOFF", and "Reset", a far cry from what other BMC provide. It also requires another Mac in the same LAN to be the controller, and you cant use static IPs (which is pretty common for BMC VLANs).

Also it seems unsupported for the Mac Pro (2023) this thread is about?

In addition to the problems other people have pointed out, it just doesn't have comparable specs to something like Ampere's Altra or AMD's Epyc.

24 CPU cores and 192 GB of memory might be enough for a small company's single on-site server, but those will likely want lots of storage and not want to deal with an exotic platform. Anyone who's willing to deal with ARM and macOS on a server will likely want a lot more performance per 1U.

Agreed. Rack density is the biggest thing once you get away from "a handful of servers", be it CPU, memory, or storage density.

Side note: I'm curious where this 1:8 ratio of cores:memory came from. Everyone seems to be standardising on it, even at VM levels

SEVEN AND A HALF THOUSAND DOLLARS? No wonder we do all our Mac CI work on Mac minis.

Macs remain completely horrendous to remote-administer.

Bro, come on. Apple hasn't even tried to pretend they're in the server game for at least a decade now. You're paying for dozens of features you'll never use. They don't even have drivers for their own hardware outside of fucking windows, let alone documentation about how their machines are intended to work.

I love my macs but apple will never be a cost or reliability solution to computation unless they do a complete 180. They're a boutique/luxury firm that can choose which markets they want to dominate. If you think there's not a small horde of people at Apple trying to figure out how to pitch bringing the xserve back, I have a bridge to sell you. But unless they can convince CTOs to become sysadmins it'll never be a viable way for a company as large as Apple to drive hardware development.

I think they could be a decent provider for SOHO options... such as the space filled by Synology on the lower end. It's pretty competitive on cost, but Apple could really play it in concert with their wifi/airplay solutions etc. Like a home-base for SOHO.

Then again, the more I think on it, the worse an idea it is. Too locked-in and out for being able to run the likes of Jellyfin, etc. It could be a good option for a "Server" mac.. but would still make more sense for it to run Linux as a base.

There's no mention of ECC memory.

Apple products are not server grade equipment.

24 cores isn't that much, and you end up with a huge hulking GPU just sitting there doing nothing, you can't run a well-supported server OS on it so it can't really be used as a server, and it's extremely expensive for what it is. I'm not really sure what makes you think it's worth bringing up.

That's a rack-mounted workstation, not a server.

I use their servers in a couple of cloud environments. That little things are snappy, for what I have given them as tasks. Wish them best of luck to the future.

The same goes for most recent/powerful mini pc options. Been using an R9-5900X based one for a few years now, runs great as a home server.

Reply on news.ycombinator.com