NVIDIA Announces Grace CPU For ARM-Based AI/HPC Processor

[coder]

What would really kill X86 fast is if you could pop an ARM chip into your existing 1151 pin Gigabyte Z390 Gaming X motherboard or your Asus ROG Zenith Extreme Alpha X399 motherboard like in the socket 7 days. With a simple BIOS flash, and ARM just automagically works on what you already have. I welcome the day when AMD makes AM4 ARM chips. But I don't see them yet.

AMD already made an ARM CPU that's socket-compatible with x86, but it's not AM4:

https://www.amd.com/en/amd-opteron-a1100

[numacross]

AMD already made an ARM CPU that's socket-compatible with x86, but it's not AM4:

https://www.amd.com/en/amd-opteron-a1100

The A1100 is not socketable, it's BGA

[ezst036]

AMD already made an ARM CPU that's socket-compatible with x86, but it's not AM4:

https://www.amd.com/en/amd-opteron-a1100

I'm looking forward to them making it AM4 compatible.

[coder]

ARM doesn't give anything to you, except avoiding duopoly of AMD-Intel. Why? Because internally all CPUs are RISC in nature.

There are knowledgeble folks who would sincerely disagree on the latter point, as well as whether ARM is really even RISC. Rather than take a position on that, I just want to point out that the advantages of AArch64 include:

• simpler ISA -> simpler, more energy-efficient decoder
• fixed-sized instruction word -> wider front-end
• larger GP register file -> less spilling
• relaxed memory-consistency -> greater instruction-reordering flexibility

These are undeniable, though you can certainly debate the impact each has on performance.

[coder]

Unless the transition from x86 to ARM on desktop/workstation is smooth, almost nobody currently depending on x86 apps will be willing to make an ARM CPU their main desktop/workstation CPU.

With emulation, transitioning from x86 -> ARM is even easier than Windows -> Linux (yes, even taking into account Wine).

The thing is that anyone who's spent time with a Raspberry Pi v4 or ARM-powered chromebook knows it's a viable platform. If you're not also changing your environment, then it's probably not such a big deal for most.

BTW, Apple has successfully transitioned ISAs 3 times, if you include their current move.

[Jumbotron]

" The new SVE features for the V1 Neoverse core draw on Arm’s experience working with Fujitsu on the A64FX platform, the SVE-enabled processor that is at the heart of Fugaku, the world’s top-ranked and first #1 Arm-based supercomputer. "

" Silicon partners will have full control over SVE voltage and frequency transitions, said Bergey. This enables them to run at full frequency while executing SVE code, as Fujitsu’s A64FX CPU is able to do. "

https://www.hpcwire.com/2020/09/22/arm-targets-hpc-with-new-neoverse-platforms/

The Grace SoC will be V1 (Zeus) as N2 (Perseus) will be more for massive scale out and for TDP restricted use cases as Edge compute nodes. It could be Poseidon (V2 / N3) but ARM states that won't be out until late 2022 or 2023. Grace has to be engineered, tested and verified before then. So...V1 it is.

You know...after reading Michael's article again and the press briefing from Nvidia AND looking over ARM's own roadmaps again, I'm changing my opinion.

Nvidia's ARM based "Grace" SoC will PROBABLY indeed be built on ARM's upcoming Poseidon. The uplift Nvidia is touting seems to fit well with ARM's boilerplate about Poseidon's performance uplift of Poseidon over its Neoverse V1 and N2 brethren.

I first surmised that Nvidia would take an established ARM platform like Neoverse V1 and expand on it with the approriate NVLink memory hooks and other system tweaks like Apple did with its M1, which actually casued ARM arch versions to bump up to 8.6-A. I initially thought that a similar thing would happen with Nvidia and Grace.

Now I am convinced that Nvidia will not only build Grace on the upcoming Poseidon platform but it will also be ARM v9-A.

[coder]

Our computers are already hybrids. SSDs have own processor to manage data. GPU is own kind of processor with own principles. Samsung I think was supposed to make a prototype RAM with some sort of logic as well implemented in it for simpler computing. Heck even AMD and Intel use own "inner" processor for security workloads, and it is ARM architecture.

Your computers are already hybrids.

I think that's not what they meant. I imagine they had in mind that one OS kernel should be managing hybrid ISA CPU cores, which share a global pool of RAM. This would be an interesting project, but I'm not sure we really have anything like it, today.

[Jumbotron]

If you all want to learn more about ARM v9 and ARM's "Total Compute Vision" this should be a good start. Lots of video chats with ARM partners including a "fireside" chat with Microsoft exec.

https://www.arm.com/campaigns/arm-vision

[coder]

Clock for clock performance is not telling the full story. And not running them higher may not be just about efficiency, but also simple stability. you can get great IPC in a processor but be unable to clock it high because of the design.

I think what you're dancing around is the critical path length of the circuitry. If a core is designed for lower clock speed, its critical path can be longer, which enables more complex pipeline stages and in turn benefits IPC. However, you can't then take that exact design and crank up the clock speed. Likewise, if a core is designed to clock higher, this will naturally come at the expense of some IPC.

In general, the argument for lower clock speeds is energy-efficiency, which is why mobile and server cores tend to clock lower than desktop CPUs. However, the designers can also take advantage of that to imbue them with more IPC, depending on the silicon & power budget.

ARM has one inherent benefit over x86, in that you can scale up its front-end wider, due to its fixed-length instruction encoding. So, if one is willing to devote the silicon necessary, it's not surprising to see ARM cores that exceed x86 in IPC. And this is independent of clock speed, in which case it should even be possible to build an ARM core that clocks comparable to x86 and offers more IPC. There's just not as much incentive for it, given that ARM can beat x86 in single-thread performance with higher IPC at lower clock speeds.

[coder]

Actually this advantage is also not that clear, if we take (potential) vendor locked-in into consideration.
We may switch Xeon with Epyc and enjoy improved performance/dollar, but when we switched one ARM from another one, there is no guarantee that they share the shame extensions and have similar performance behavior.

ARM is very strict about licensees not adding their own instructions. So, in that respect, it's less susceptible to lock-in than x86. And if you adopt software that relies on a certain ISA level, you should only do so with the knowledge that you're restricting yourself to fewer CPUs.

RISC-V is the worst, though. Basically, a RISC-V CPU can add whatever the heck it wants!