• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

Nvidia Grace CPU - Nvidia enters the CPU market

Soldato
Joined
6 Oct 2009
Posts
4,040
Location
London
Nvidia has officially announced its first dedicated CPU: Nvidia Grace Superchip.

It's a server chip designed to compete against Intel Xeon, AMD EPYC and Ampere Altra.
  • 144 cores
  • ARM ISA
  • Up to 1TB DDR5
  • 5nm
  • PCIe 5
  • 1 TB/s memory bandwidth
  • Two chips connected together through a NVLink-C2C
  • 500w incl. memory
  • Ships in 2023
Nvidia did not disclose which ARM core they are using, but it is very likely either the ARM Neoverse N2 or V1. It most definitely is not a custom microarchitecture.

Nvidia claims performance is 720 on the SPECrate_2017int_base, which is the gold standard general test for a server chip. For comparison, a single 64-core Zen 3 AMD EPYC 7773X scores about 475 (which is a world record for a 1-node-1-CPU config). And dual socket ones score in the 850 range. So slightly behind two AMD EPYC 7773Xs if we are to believe Nvidia's claims.

gS3SUMw.png

https://nvidianews.nvidia.com/news/nvidia-introduces-grace-cpu-superchip
https://www.tomshardware.com/news/n...laims-arm-chip-15x-faster-than-amds-epyc-rome
 
Last edited:
What's the TDP of dual chip EPYC? Is performance per watt in Nvidia or AMD's favour?

The EPYCs would be 280w each, and that's before Ram. Given that Ampere Altra Max (128-core, Neoverse N1, TSMC 7nm) did beat Milan EPYCs in perf/watt, I would expect this will do as well.
 
Interesting product. But what will be more interesting is that it will be up against Zen 4 Genoa in 2023 which is when this will be out.

NVidia are probably just pleased they won't be paying to use AMD CPUs anymore.
 
Meh....

Interesting product. But what will be more interesting is that it will be up against Zen 4 Genoa in 2023 which is when this will be out.

NVidia are probably just pleased they won't be paying to use AMD CPUs anymore.

Yeah.

Pretty meaningless to compare an unreleased chip to current products, its good for headline garbing and looks good if you ignore its competitors.

With 144 cores its slower that a similar current gen EPYC dual socket.

Genoa, 96 Zen 4 cores per socket will launch later this year, before this, so before Nvidia can get this out a dual socket Genoa will be around 2X as fast

In H1 2023 this will launch along side Bergamo, 128 Zen 4 cores per socket, in dual socket form Bergamo will be 3X as fast.

You know what AMD should do? Just join in with these stupid games, just hold Bergamo up in front of the camera, like she already did with Genoa but this time say laud and clear "lanching at the same time as Nvidia Grace and 3 times as fast" Boom... we can play these stupid games too Nvidia.
 
This isn't about who's better or worse (we need to wait for a whole year before we can even assess it), or what metrics to use (perf/rack space, perf/watt, absolute perf, etc). We'll get to these when this is released, I don't expect a first-gen Nvidia product to beat Intel and AMD (well, they may beat Intel but not AMD).

More competition is not a bad thing. CPU market duopoly is vanishing very fast and that's a very good thing.
 
This isn't about who's better or worse (we need to wait for a whole year before we can even assess it), or what metrics to use (perf/rack space, perf/watt, absolute perf, etc). We'll get to these when this is released, I don't expect a first-gen Nvidia product to beat Intel and AMD (well, they may beat Intel but not AMD).

More competition is not a bad thing. CPU market duopoly is vanishing very fast and that's a very good thing.

This will have no effect on the likes of us anyway, for consumer Desktop Nvidia would need an X86 licence from Intel and an AMD64 licence from AMD.
 
This will have no effect on the likes of us anyway, for consumer Desktop Nvidia would need an X86 licence from Intel and an AMD64 licence from AMD.

That assumes Nvidia would even have any interest in that, which they wouldn't. They're not going to even want to enter the x86 market. They wanted to buy ARM.

This wouldn't be Nvidia making a CPU for you to DIY build, they'll likely do the Apple approach, they'll build the entire SoC and memory package, and use their partners to bundle that into laptop/desktop form factors.
 
That assumes Nvidia would even have any interest in that, which they wouldn't. They're not going to even want to enter the x86 market. They wanted to buy ARM.

This wouldn't be Nvidia making a CPU for you to DIY build, they'll likely do the Apple approach, they'll build the entire SoC and memory package, and use their partners to bundle that into laptop/desktop form factors.


Nvidia in practical terms can't enter x86 anyway because those who hold the license will charge so much for it that Nvidia wouldn't be able to compete on price with Intel and amd. X86 due to that license is uncompetitive
 
Nvidia in practical terms can't enter x86 anyway because those who hold the license will charge so much for it that Nvidia wouldn't be able to compete on price with Intel and amd. X86 due to that license is uncompetitive

Pretty much, and personal computing isn't purely x86 anymore. With smartphones, tablets and now laptops and desktops on ARM, and ARM's big push on the server side of things, x86 itself has serious competition for the first time in maybe 20 years.

Qualcomm (after buying Nuvia) have confirmed they're entering this market, as well as Ampere (with their upcoming custom microarchitecture), and quite likely Nvidia too. Microsoft themselves are heavily investing into making their own ARM chips as well to be competitive against Apple in the business laptop business in terms of performance and battery life (surface line).

Everything has changed in the last 18 months or so. Every major linux distro has a very stable ARM version now, almost every Linux package manager now has access to ARM builds and over 99% are already native. This wasn't the case 18 months ago. Software side of things is pushing ahead at lightning speed. I don't expect there to be any difference between ARM and x86 in 2 years in terms of software availability/compatibility, with the one exception being gaming which will take longer to adapt.
 
Slap in a PCIe x16 slot or two, cut it to one CPU with fewer cores and it could be a great gaming PC. I wonder if Nvidia are going after the next console contract?
 
Slap in a PCIe x16 slot or two, cut it to one CPU with fewer cores and it could be a great gaming PC. I wonder if Nvidia are going after the next console contract?

They couldn't really compete in the last two generations, Microsoft and Sony wanted a single SoC for CPU/GPU which Nvidia couldn't reasonably provide, so that left AMD as the only real contender. With new consoles not coming until 2025 at the easiest, Nvidia can be in a very good position to have a go at it.
 
Back
Top Bottom