I have an idea for Apple's next great product. Bring back Xserves, but with M-series processors in them this time. Then hire the Asahi Linux developers and get Linux running very solidly on them. Make it so you can buy them with either macOS or Linux pre-installed.
Sincerely,
lockhouse
Not having to reverse engineer drivers for pretty much every single component would probably require a fraction of the effort though. And it's not like Apple would really need to actually maintain their own distro, releasing the drivers (which they are more than capabable of doing) would be perfectly sufficient.
What does any of this have to do with Neoverse?
The biggest gap right now (afaik) is Apple's virtualization limits. Regardless of the cores in your system, any Mac can host no more than 2 instances of MacOS. It's a huge kneecap to how Macs can be used as a server, and a death-blow to any serious hosting company that wants to provision them as a virtual resource.
> I would think the performance per watt would be very competitive with other server vendors on the market as well.
I can go get a 4 core Ampere ARM instance for free. The market goes so much cheaper than Apple's CPUs it's not even funny. Performance-wise, Apple kinda got their lunch ate by Nvidia's Grace server. There isn't really a demand to fill for high-performance ARM computing, and the cheap end goes so cheap that Apple can't even touch them.
I see two ways for Apple to escape this:
1. They can double-down on their ecosystem investments to create a "CUDA-killer" and accompanying hardware, spending tens-of-billions of dollars across multiple years to obsolete Nvidia's industry ties.
2. Initiate or join an Open Source pact to write a CUDA replacement or translator a-la Vulkan/DXVK that gives vendors no choice but to abandon Nvidia.
Of course, the easier play is to just pretend all of this is beneath you and continue building your current ecosystem. It's what I would do, and I don't see a rational reason why Tim Cook would read your pitch and greenlight a project like that.
Fairly good deal, the only problem being that it's Oracle.
ps. if you do decide to give them a try, be sure to upgrade to "pay-as-you-go" (and never go beyond the free offering) otherwise they will close your account without any warning (ask me how I know).
That being said, I wouldn't trust the server as far as I could throw it. It's a bad deal unless you have ephemeral compute requirements that don't matter too much.
Why not? Even just Apple's internal expense could justify it - for the same reason Google, Microsoft, Amazon etc all dog food their own cloud. Apple currently pays AWS, Akamai, Cloudflare, etc lots of $$$.
As the current focus is to expand on services and managed offerings you would focus on controlling such a cost (as is typical Apple).
> I can go get a 4 core Ampere ARM instance for free. The market goes so much cheaper than Apple's CPUs it's not even funny.
Is the market full of people that only use 4 cores? If that was really the case AWS would have closed shop already instead of being the market leader. It's not like Oracle doesn't charge you for anything else.
> Apple kinda got their lunch ate by Nvidia's Grace server
And where is Grace? Which cloud currently? It doesn't work that way. If all you think about a server and its ecosystem is hardware then lots of Android phones have better hardware than iphone e.g. in screens, battery life, etc. Was that really the deciding factor? No.
I mean by your logic, AMD's EPYC has been winning Intel Xeon's by years. Has Xeon been decimated? No. There's still room for both. AMD can't even produce enough. Considering Apple can exclusive lock TSMC down for a year on bleeding edge, who can really compete if Apple was in?
Ampere CPUs on Oracle, Hetzner and (I assume?) Azure are still remarkable cheap compared to Intel/AMD if they are sufficient for your workflows.
I'm not sure if the margins in this market are high enough for Apple to enter it and they would probably still have issues scaling production capacity if they are serious about it.
Hetzner has always been there. It didn't impact Intel/AMD use cases. There are lots of cheap providers (since the start of time) and I doubt that's the space Apple would ever get into.
So... who would they be selling to, now? I genuinely think such a "server Mac" would echo the Xserve without a fundamental shift in the way Apple treats the software community.
Why? And why is it only about hardware? So AWS, GCP and Azure all run Intel/AMD, who do they need to dethrone?
For all it matters Apple "cloud" could even use these nVidia servers themselves. Not like AWS only uses Graviton. So, so?
> So... who would they be selling to, now? I genuinely think such a "server Mac" would echo the Xserve without a fundamental shift in the way Apple treats the software community.
Right, why would they only do hardware? As is typical Apple the $ is in the ecosystem and services. The competition would be to offer cloud services that has a better experience just like they do for phones and others. The macbook pro was always more expensive and yet people still purchase them.
You are just back around to describing Xserve again. Regardless of whether it is software or hardware-based, an Apple Cloud has no upsell in today's market. That's why Xserve died, and it's why it would die again.
Which is both less relevant and quite different for enterprise services than consumer products. It's also not that obvious that Apple has a lot of expertise in this area and could provide something better than AWS/Azure/etc.. and Apple is not exactly the company which is willing to compete on lower margins.
I would actually argue that is the easy part. Apple already has access to cutting-edge nodes if they want to pay for them. The hard part is designing a server chip from Apple Silicon, a decidedly mobile SOC.
For instance; what if Apple wanted to ship a system that competes with a DGX GH200? Where would you even start with something like the M2 Ultra?
- You would either need more cores per SOC, or more SOCs per chipset to match a DGX system
- Custom memory controllers are a must (running them at regular speeds depends on how much bandwidth is left)
- PCI would need to exposed on-hardware and supported in MacOS, Linux and third-party software
- An multi-instance interconnect like Infiniband would need to be working and attractive (as well as a CUDA translator, if they want any hope)
And that's just what they would need to match the high-performance ARM industry right now. If they wanted to compete or knock Nvidia's block off, it would take a lot more investment, time and attitude-changing.
Just on par with AWS graviton...
Likewise Apple also had their own Linux distribution that went nowhere, MkLinux.
MkLinux (MicroKernel Linux) was not a distribution but a modified Linux kernel to be loaded as a Mach server to run under CMU (later OSF) Mach 3. It abstracted idiosyncrasises of the Apple NuBus and peripherals away from the Linux kernel that did not (or could not altogether) support – the NuBus itself and the NuBus attached devices.
The distribution was a stock RedHat v4 or v5.
Amusingly, despite CMU Mach 3 being a µ-kernel, there was nothing «micro» about the size of its own binary, and – off the top of my head – the Mach binary size was larger than that of the vmlinuz at the time.
If anything, it was actually pity it didn't took off, as its design was quite interesting.
There are?
Especially after being selected for EU super computer effort.
Isn't it years behind ARM/x86 performance wise?
> nonsense to dismiss it as only being for embedded
They unlikely to be competitive anywhere else without significant investment.
> but it can also scale up to "real" compute
Did anyone doubt that you could theoretically scale up ARM to be competitive with x86 back in ~2000? It still took them several decades to achieve that..
I don't expect them to open source their processors, but they do definitely contribute back to the community. E.g. they open sourced their first risc-v vector implementation (ocelot) on top of BOOM. It's in order and they used it as a reference for their later out of order commercial designs, but they also could've keep it to them self's.
RISC-V openness has nothing to do with processor implementation, it's solely in the instruction set.
The advantage it has is that it provides compatibility across vendors and anyone can become a vendor of RISC-V IP. Comercial IP doesn't affect that, and if anything makes multiple IP vendors more likely.
The availability of server processors depends on the market rather than any technical limitation. If someone decides to take the risk of investing in the development of a strong server product then we'll get it. It depends on some company believing there is a market for it.
Just because there is a new ISA does not mean a company cannot design a modern CPU.
For example, the Vision Star Five is uses an in order CPU core that is competitive with the out of order A72 core. The performance gap on day one is much smaller. It also has UEFI support and opensource GPU drivers for Power VR are in development by imagination, so the software support will also quickly catch up or surpass ARM SBCs.
The same will happen in the server market. Nobody who works on a RISCV server chip has to wait for ARM to deliver a CPU core that doesn't suck.
Or that they show up in computer shops for regular buyers, instead of computer enthusiasts, used to install GNU/Linux into any kind of motherboard with a CPU on it.
Cause why not.. having more architectures (for no apparent reasons) sounds cool.