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

[drakonas777]

I'd like to correct myself a bit. I'm not against ARM in general. I just can't see any attractive performant and affordable "ARM desktop board". Yes, there are some pretty cool ARM boars here and there, like that Xavier, but they are more like a beefed up versions of "classical ARM embedded systems" than a proper desktop boards. I mean there are always some SOM/daughter/mother board bullshit there, or some effing eMMC, or limited peripherals/buses or something. They "smell like" a simple SoC dev boards. What I want is something like those server ARM SoC developer boards - proper ITX/ATX format, proper ports/interfaces, high performance ARM core architecture (not mobile shit). But these are multiple thousands of dollars however. I want them (or something similar) at competitive price with mainstream x86 desktop. When this happens I'll agree with "age of ARM" meme (I mean specifically Windows/Linux ecosystem, not Apple HW). I know that you can get somewhat "desktopish" experience with some ARM SBCs, but I had a full blown ARM desktop platform in mind at the competitive price.

BTW, let's be honest here: in order to be attractive ARM should offer even more performance at the same price to compensate all ISA transition-related issues. They do it in enterprise, I want it for consumer too.

[Dawn]

• The objective difference/distance between x86 and ARM is most likely less than 10%
  • In other words: Any potential performance advantage of the NVIDIA's ARM CPU over x86 CPUs is completely unrelated to x86-vs-ARM
• In year 2023, new x86 CPUs will most likely have 10-20% higher instructions-per-clock than Zen 3 and Rocket Lake

I agree with you, actually - the "x86 tax" is an overblown concern. There's an underconsidered factor, IMO, which is impact of a complex ISA (and platform) on design and validation time and cost, but anyone who claims x86 is vastly more efficient/faster/whatever than ARM (or vice versa) because of RISC vs CISC or something is clearly not someone who has spent a significant time doing microarchitecture work.

That being said, ARM's Austin design group has been consistently executing well. Today's Neoverse cores already have iso-clock performance that compares favorably to Intel and AMD cores. N2 is another 40% on top of that, V1 50% (but at lower efficiency); I fully expect that in early 2023, the cores Nvidia is using will be competitive with what's emanating from the x86 vendors. They've had a lot of big generational gains in a row, essentially since the A75, and that does add up.

[qarium]

I'd like to correct myself a bit. I'm not against ARM in general. I just can't see any attractive performant and affordable "ARM desktop board". Yes, there are some pretty cool ARM boars here and there, like that Xavier, but they are more like a beefed up versions of "classical ARM embedded systems" than a proper desktop boards. I mean there are always some SOM/daughter/mother board bullshit there, or some effing eMMC, or limited peripherals/buses or something. They "smell like" a simple SoC dev boards. What I want is something like those server ARM SoC developer boards - proper ITX/ATX format, proper ports/interfaces, high performance ARM core architecture (not mobile shit). But these are multiple thousands of dollars however. I want them (or something similar) at competitive price with mainstream x86 desktop. When this happens I'll agree with "age of ARM" meme (I mean specifically Windows/Linux ecosystem, not Apple HW). I know that you can get somewhat "desktopish" experience with some ARM SBCs, but I had a full blown ARM desktop platform in mind at the competitive price.
BTW, let's be honest here: in order to be attractive ARM should offer even more performance at the same price to compensate all ISA transition-related issues. They do it in enterprise, I want it for consumer too.

ARM is not on Desktop or Workstation YET....

ARM right now is only in mobile market and server market.

but on server market a 80-128core ARM cpu is at 4000€ and has similar performance to a 40core 10nm 8000€ Xeon.

by this alone you can see ARM is the future... a 16core AMD 2950X is at ~700€ yes you need 32cores for ARM to be able to compete but it could only cost you ~400€

the X86_64 Tax is that high.

[Dawn]

ARM is not on Desktop or Workstation YET....

ARM right now is only in mobile market and server market.

but on server market a 80-128core ARM cpu is at 4000€ and has similar performance to a 40core 10nm 8000€ Xeon.

by this alone you can see ARM is the future... a 16core AMD 2950X is at ~700€ yes you need 32cores for ARM to be able to compete but it could only cost you ~400€

the X86_64 Tax is that high.

The x86 tax, as commonly interpreted ("in an inherent performance or efficiency disadvantage", rather than "something that may increase design time and cost"), is minor at best.

[piotrj3]

Yet I sit here running an M1 based Mac Book Air that is as good as my high end desktop and runs all the software I need to run on a laptop. The future is already here, it just requires people to swallow the right pill and wake up from their coma.

ARM doesn't give anything to you, except avoiding duopoly of AMD-Intel. Why? Because internally all CPUs are RISC in nature. M1 of apple is actually not more energy efficient in multithread workloads then AMD stuff in laptops. Why? Because desktop computers still need great single threaded performance. And single threaded performance costs a lot of energy. When M1 is really impressive by memory controller and memory speed and fact it can match AMD and Intel, M1 4 performance cores for 25W TDP in laptop like chassis is not impressive, it is actually underwhelming (comparing to for example 4700U that has 8 cores 16 threads also in max 25W package) and is actually outdated design (by date of premiere we should compare Zen3 mobile designs and M1).

ARM is not saving any of you in desktop computers and will not outperform intel/amd in your everyday workloads. On server, or in well scalable HPC enviroment all single thread performance bullshit is not needed, just make cores simple low power and because they are simple and small you can make tons of them.

[Dawn]

ARM doesn't give anything to you, except avoiding duopoly of AMD-Intel. Why? Because internally all CPUs are RISC in nature. M1 of apple is actually not more energy efficient in multithread workloads then AMD stuff in laptops. Why? Because desktop computers still need great single threaded performance. And single threaded performance costs a lot of energy. When M1 is really impressive by memory controller and memory speed and fact it can match AMD and Intel, M1 4 performance cores for 25W TDP in laptop like chassis is not impressive, it is actually underwhelming (comparing to for example 4700U that has 8 cores 16 threads also in max 25W package) and is actually outdated design (by date of premiere we should compare Zen3 mobile designs and M1).

ARM is not saving any of you in desktop computers and will not outperform intel/amd in your everyday workloads. On server, or in well scalable HPC enviroment all single thread performance bullshit is not needed, just make cores simple low power and because they are simple and small you can make tons of them.

V1 and friends are a long way from "simple and small" cores. V1 is, in many respects, wider than Zen3 (can sustain execution of 8 ops on mop cache hit, for instance, vs 6 from Zen3.) It is a big, complex machine.

[andre30correia]

a good arm is what we need but in desktop something open not like apple produts

[polarathene]

Yet I sit here running an M1 based Mac Book Air that is as good as my high end desktop and runs all the software I need to run on a laptop. The future is already here, it just requires people to swallow the right pill and wake up from their coma.

Can't run CUDA even with eGPU support since Apple blacklists NVidia. Other OS support is limited, macOS itself (ARM aside) is limited, notably in official support. I've heard bad things about some software compatibility for doing dev with, but I haven't been following so maybe they've addressed that, how well does Docker run on it btw? Did they address the SSD issue where it was writing massive amounts daily on swap?

I use 32GB RAM and often in the 20GB+ range, I need a good GPU that can do compute work especially with proprietary software that only supports CUDA (no alternatives, proprietary or open-source are anywhere near parity in features or performance, it's a niche industry). Work machines use 128GB RAM, lots of fast NVMe storage, leverage lots of PCIe. The future ain't here yet for everyone, but Grace sounds like it could be sweet.

M1 is a good example of progress, but it's still not quite there yet.

[polarathene]

by this alone you can see ARM is the future... a 16core AMD 2950X is at ~700€ yes you need 32cores for ARM to be able to compete but it could only cost you ~400€

the X86_64 Tax is that high.

So what you're saying is that if I have workloads that are single threaded because they cannot be designed to be multi-threaded, ARM is still a good while away from being competitive? Bring on the hybrid x86 + ARM systems then :P

[qarium]

The x86 tax, as commonly interpreted ("in an inherent performance or efficiency disadvantage", rather than "something that may increase design time and cost"), is minor at best.

tell me what is more important how the X86 Tax is messured or the point that the tax exist ?

i really don't care how it is messured in price or in efficiency or in performance....

for me it is only relevant that the tax exist.