Question Nvidia to enter the server CPU market

[NTMBK]

NVIDIA Unveils Grace: A High-Performance Arm Server CPU For Use In Big AI Systems

www.anandtech.com

Based on Neoverse cores, not a Denver derivative.

This makes sense with the ARM acquisition attempt. They want to fully control the server stack.

 

[A///]

Isn't Grace just a bog standard ARM design?

Customised cores, IIRC. I haven't followed it much myself because I have little interest in things that cost as much as a luxury vehicle.

 

[Saylick]

Isn't Grace just a bog standard ARM design?

I think they are? The IO and fabric are Nvidia's design though.

 

[igor_kavinski]

The graph (apparently using some odd normalization and rounding) is only talking about performance and throughput, both saying nothing about power efficiency at all

Grace is giving more throughput than Genoa when both are used in a 5MW data center. So obviously Grace is more power efficient. Or am I missing something?

 

[DrMrLordX]

Customised cores, IIRC. I haven't followed it much myself because I have little interest in things that cost as much as a luxury vehicle.

I think they are? The IO and fabric are Nvidia's design though.

From what I've read it's Neoverse V2 cores with custom IO and fabric. But I wasn't sure if that was still true.

 

[Markfw]

Grace is giving more throughput than Genoa when both are used in a 5MW data center. So obviously Grace is more power efficient. Or am I missing something?

When an independent source benchmarks these then we can talk. Until then its all marketing.

 

[NTMBK]

Wide Horizons: NVIDIA Keynote Points Way to Further AI Advances

Chief Scientist Bill Dally described research poised to take machine learning to the next level.

blogs.nvidia.com

View attachment 85172

AMD may have to go ARM or RISC-V to compete on power efficiency.

They're comparing systems with DDR5 DIMMs against a system with LPDDR5X, 5nm against 4nm process, and traditional dual-socket against tightly integrated pairs of CPUs on the same package. It's hardly an apples to apples comparison of x86 Vs ARM!

 

[Doug S]

They're comparing systems with DDR5 DIMMs against a system with LPDDR5X, 5nm against 4nm process, and traditional dual-socket against tightly integrated pairs of CPUs on the same package. It's hardly an apples to apples comparison of x86 Vs ARM!

Yes they are comparing SYSTEMS. If one is designed to use DIMMs and the other LPDDR then that's how you should compare them. If you were comparing a PC laptop with a Mac laptop would you say "it is unfair comparing performance because Apple uses LPDDR" but think it is fine comparing expandability? I agree it is not a valid comparison of x86 and ARM but it is a valid comparison between two systems, which were made differently as their respective OEMs made different decisions along the way.

And N5 vs N4 is pretty minor. If it was N7/N6 vs N5/N4 that might be more valid but if both systems were shipping at the time of the comparison then how would be it be the fault of the one using N4 that the other was using N6? The tradeoff for that decision leading to worse performance would be lower cost to make the chips, same as the tradeoff between higher memory bandwidth vs ability to expand memory in DIMMs vs LPDDR.

 

[A///]

When an independent source benchmarks these then we can talk. Until then its all marketing.

Would nvidia legal put a stop to that? The only vague benchmarks I can think of off the top of my head is Jim Keller's strange blurry graph scale where he set zen 5 way ahead of the grace hopper platform.

 

[NTMBK]

Yes they are comparing SYSTEMS. If one is designed to use DIMMs and the other LPDDR then that's how you should compare them. If you were comparing a PC laptop with a Mac laptop would you say "it is unfair comparing performance because Apple uses LPDDR" but think it is fine comparing expandability? I agree it is not a valid comparison of x86 and ARM but it is a valid comparison between two systems, which were made differently as their respective OEMs made different decisions along the way.

And N5 vs N4 is pretty minor. If it was N7/N6 vs N5/N4 that might be more valid but if both systems were shipping at the time of the comparison then how would be it be the fault of the one using N4 that the other was using N6? The tradeoff for that decision leading to worse performance would be lower cost to make the chips, same as the tradeoff between higher memory bandwidth vs ability to expand memory in DIMMs vs LPDDR.

Oh for sure, it's totally valid to compare systems. I'm not saying it's a bogus comparison, it's just silly to use that to make an ARM Vs x86 point, there are much bigger system-architecture level factors at play.