Nvidia Grace falls short of Threadripper 7000 in head-to-head Linux benchmarks

Nvidia GH200 SC23 Announcement
(Image credit: Nvidia)

A 39-test comparison between Nvidia's Grace server CPU and AMD's Threadripper 7980X and Threadripper Pro 7995WX shows that Nvidia's Arm-based chip isn't far behind AMD's powerhouse. The benchmarks come courtesy of Phoronix, which put the three CPUs to the test in various Linux applications, and while it's clear that Grace isn't a champion, it's still impressive.

In the comparison, Phoronix tested a workstation from GPTshop.ai equipped with a Grace-Hopper GH200, which has both a Grace CPU and a Hopper-based H200 GPU. The Grace CPU has 72 cores, 480GB of LPDDR5X memory, and is based on the Arm architecture. Individual Grace chips aren't sold alone; they're only found in CPU-GPU combo devices like GH200 and in the double-chip Grace Superchip. By contrast, Threadripper 7000 features multiple models and can come with up to 1TB of DDR5.

While there were 39 individual tests, Phoronix did not include any overall metric of performance to measure how well Grace did against the 64-core 7980X and 96-core 7995WX. However, we tallied up the wins for each CPU, and ultimately, Grace beat the 7980X in 17 tests and the 7995WX in 15. That means Grace lost more tests than it won, and while this doesn't account for overall performance, we suspect Grace would also be behind. Phoronix also tested the same Grace workstation against AMD's Epyc server CPUs.

This isn't a great result for Grace, but it's not all that bad under the circumstances. Threadripper 7000 enjoys many advantages: many apps are optimized solely for x86 and not Arm, Threadripper 7000 has much more aggressive clock speeds than the efficiency-focused Grace chip and far more L3 cache (7980X has more than double, and the 7995WX more than triple).

The comparison here might not be entirely apples to apples. After all, Grace is designed for servers and is focused on efficiency, while Threadripper is a bona fide workstation chip that pulls out all the stops for maximum performance. Nvidia hasn't disclosed the TDP of a single Grace CPU, and Phoronix didn't have power data to publish. However, the Grace Superchip has a TDP of 500 watts, implying just one Grace has a TDP of 250 watts or more. For comparison, the 7980X and 7995WX are rated for 350 watts, which may mean Grace is more efficient.

Grace also compares favorably to Intel's Sapphire Rapids in terms of efficiency. Benchmarks from earlier this month show Nvidia's server CPU losing to Intel's in raw performance but indicate that the tables are turned regarding efficiency

Matthew Connatser

Matthew Connatser is a freelancing writer for Tom's Hardware US. He writes articles about CPUs, GPUs, SSDs, and computers in general.

  • Notton
    I'm disappointed that it doesn't have any power consumption and efficiency numbers to go with the benchmark numbers.
    Like okay, they can't get the numbers to read from the mobo, but surely they could measure it at the wall.

    Surely the ARM chip is more efficient... right? If not, that'd be funny too.
    Reply
  • JTWrenn
    For what these are used for perf per watt is king. If it is close to similar in power but 30% lower on power, that is a big win. ie 250 tdp vs 350 tdp. Hard to say for sure without a full blown hands on test with power numbers but to me this sounds like a decent win especially for a first of it's line chip. Next round should be very interesting especially with all the capital Nvidia has to play with now.
    Reply
  • thisisaname
    Given it is a new chip could there also be room for optimisation of the software?
    Reply
  • HaninTH
    Is this ARM's chance to make a major move to displace x86 now that it has a killer app to drive it (AI & ML)? Will that be enough to get developers to code more for ARM vs x86 and cause the cascade to occur, at least in the server space?

    I still do not believe ARM will make major inroads in the desktop, despite Apple, as the majority of software is still made for x86. The eco system that x86 currently has will be hard for ARM to match in the near term. They don't have the same flexibility that x86 hardware currently enjoys.

    Guess we'll have the next 5 or so years to see how this turns out!
    Reply
  • purposelycryptic
    Notton said:
    I'm disappointed that it doesn't have any power consumption and efficiency numbers to go with the benchmark numbers.
    Like okay, they can't get the numbers to read from the mobo, but surely they could measure it at the wall.

    Surely the ARM chip is more efficient... right? If not, that'd be funny too.
    Just a guess, but if it were more efficient, we'd probably have numbers. Also, it's a workstation processor, designed for power, vs a server processor, designed for efficiency - if it isn't significantly more efficient, that would probably be pretty bad news...
    Reply
  • purposelycryptic
    JTWrenn said:
    For what these are used for perf per watt is king. If it is close to similar in power but 30% lower on power, that is a big win. ie 250 tdp vs 350 tdp. Hard to say for sure without a full blown hands on test with power numbers but to me this sounds like a decent win especially for a first of it's line chip. Next round should be very interesting especially with all the capital Nvidia has to play with now.
    For workstations, total performance is king; for servers, efficiency is king. A server processor should be using a good bit less juice than a same-generation top-end workstation processor.
    Reply
  • purposelycryptic
    HaninTH said:
    Is this ARM's chance to make a major move to displace x86 now that it has a killer app to drive it (AI & ML)? Will that be enough to get developers to code more for ARM vs x86 and cause the cascade to occur, at least in the server space?

    I still do not believe ARM will make major inroads in the desktop, despite Apple, as the majority of software is still made for x86. The eco system that x86 currently has will be hard for ARM to match in the near term. They don't have the same flexibility that x86 hardware currently enjoys.

    Guess we'll have the next 5 or so years to see how this turns out!
    I don't think anyone really wants ARM desktops - fracturing the market, compatibility issues...

    Most home users want a computer that just works, the same way their old one did.

    Most businesses really don't want to deal with compatibility issues with the part of their software stack that is centuries old and just barely still runs on Windows (and they all have at least one legacy program essential to their operations - everyone does). They also don't want to pay to have their techs retrained on all of ARMs little quirks, and the support department expanded because of millions of calls from employees whose new computers are acting just different enough to make them panic.

    Most enthusiasts and gamers don't want to deal with major incompatibilities with their older software and games.

    Developers REALLY don't want it, because they'll have to release two versions of everything for probably at least a decade or so, which hurts just to think about.

    Laptop users might think they want it, because of better battery life, but what they really mean is that they want the same exact experience that they have now, just with better battery life. At least, that's almost certainly the case for the vast majority of them, that use their laptops for work, browsing, and entertainment, with no clue and no interest in how things work actually work, or why. They still want all their software to work exactly the way it always has, no matter how ancient or obscure. They definitely don't want to have to learn anything new just because they got a new laptop.

    Apple users accepted it because they will accept anything Apple does, and because they didn't have any other option; the entire Mac ecosystem moved, all at once, all together - it was that or move to Windows, and that was never going to happen. There was no fracturing the market, because Apple IS the market, and, because they have complete control, they were able to force developers to address compatibility issues ASAP.

    None of those things are the case for PC.
    Reply
  • Notton
    I wouldn't mind having an ARM desktop/laptop, but only if x86 compatibility is completely seamless and bug free.

    As in, I don't want to notice it's not running Windows on x86-64, because nearly all of my games and programs run on that.

    And I am one who notices all the lack of programs an x86 Chromebook has.
    "It's light weight and can run on a n95 celeron with 8GB of RAM and eMMC storage". Well yeah, it literally doesn't do anything other than email and chrome. I can't even resize photos easily on it, such a hassle.
    Reply
  • Pierce2623
    Notton said:
    I'm disappointed that it doesn't have any power consumption and efficiency numbers to go with the benchmark numbers.
    Like okay, they can't get the numbers to read from the mobo, but surely they could measure it at the wall.

    Surely the ARM chip is more efficient... right? If not, that'd be funny too.
    If the ARM chip that tops out at like 3.6GHz isn’t more efficient than the x86 chip that hits 5GHz, that would be crazy.
    Reply
  • Pierce2623
    Notton said:
    I wouldn't mind having an ARM desktop/laptop, but only if x86 compatibility is completely seamless and bug free.

    As in, I don't want to notice it's not running Windows on x86-64, because nearly all of my games and programs run on that.

    And I am one who notices all the lack of programs an x86 Chromebook has.
    "It's light weight and can run on a n95 celeron with 8GB of RAM and eMMC storage". Well yeah, it literally doesn't do anything other than email and chrome. I can't even resize photos easily on it, such a hassle.
    Yeah Windows on ARM is useless until it actually runs all 64 bit Windows applications (including games)
    Reply