News Nvidia Unveils 144-core Grace CPU Superchip, Claims Arm Chip 1.5X Faster Than AMD's EPYC Rome

[Admin]

Nvidia unveiled its new 144-core Arm-based Grace CPU Superchip at GTC 2022.

Nvidia Unveils 144-core Grace CPU Superchip, Claims Arm Chip 1.5X Faster Than AMD's EPYC Rome : Read more

 

[Liquidrider]

Looks like Nvidia is more worried about AMD than they are about Intel. Can tell because they didn't compare their chips to Xeon at all. Instead focusing on AMD's 2020 release of Epyc Rome. That doesn't exactly scream homerun, but impressive none the less.

Regardless, I am noticing Intel is becoming more and more an outsider in the server CPU space even by the likes of their own competition.

 

[jeremyj_83]

Looks like Nvidia is more worried about AMD than they are about Intel. Can tell because they didn't compare their chips to Xeon at all. Instead focusing on AMD's 2020 release of Epyc Rome. That doesn't exactly scream homerun, but impressive none the less.

Regardless, I am noticing Intel is becoming more and more an outsider in the server CPU space even by the likes of their own competition.

Rome was released in August 2019.

 

[renz496]

Looks like Nvidia is more worried about AMD than they are about Intel. Can tell because they didn't compare their chips to Xeon at all. Instead focusing on AMD's 2020 release of Epyc Rome. That doesn't exactly scream homerun, but impressive none the less.

Regardless, I am noticing Intel is becoming more and more an outsider in the server CPU space even by the likes of their own competition.

because that's what they used on their DGX A100? the future DGX will use nvidia in house Grace ARM CPU.

 

[Neilbob]

All it takes is a single benchmark to be faster in order to claim the same performance margin applies to everything.

A wait-and-see approach is always needed when manufacturers start spouting numbers.

 

[spongiemaster]

because that's what they used on their DGX A100? the future DGX will use nvidia in house Grace ARM CPU.

Exactly. The A100 based DGX system uses dual 7742 Rome based Epyc CPU's. Nvidia is comparing their new CPU's to the one's they are currently using.

 

[spongiemaster]

All it takes is a single benchmark to be faster in order to claim the same performance margin applies to everything.

A wait-and-see approach is always needed when manufacturers start spouting numbers.

These DGX boxes aren't general purpose servers, they are used for pretty specific tasks. If Nvidia designed these CPU's in house, we can be pretty sure they tuned them for use cases relevant to the target market.

 

[msroadkill612]

These DGX boxes aren't general purpose servers, they are used for pretty specific tasks.

This is the elephant in the room i suspect.

Its a spoiler announcement re a 2023 future product, in a series that has lacked the impact of the holistic epyc+GPU platform, & more a reflection of how scary the new gpuS from amd are considering the failed ARM deal.

what NV can do w/ arm, amd can dobetter on Infinity Fabric mixed with Xilinx FGPA magic.

 

[Makaveli]

lol what is with both intel and nvidia cherry picking 2 year old AMD hardware to benchmark against the current offerings.

First intel using the old threadrippers and now NV using Rome and not Milan for their benchmarks.

 

[Deleted member 14196]

Because it makes them look better on paper and then people read it and say oooh ahhhh

 

[HideOut]

compairing to a pretty old chip when AMD's newer faster chips are out. Saying "early 2023" and by then intel and AMD will likely have updated. And its not X86. Fail.

 

[escksu]

Lets just wait and see.. The performance numbers were based on simulations, so this means there is no physical silicon at all.

 

[renz496]

compairing to a pretty old chip when AMD's newer faster chips are out. Saying "early 2023" and by then intel and AMD will likely have updated.

because that is the CPU that being use inside their DGX A100. the article already mention it.

Nvidia claims the Grace CPU Superchip offers 1.5X more performance in a SPEC benchmark than two of the last-gen 64-core AMD EPYC processors in its own DGX A100 servers

also nvidia are not selling this CPU in the same manner as AMD or intel selling their x86 server chip. most of this Grace CPU most likely need to be bought with nvidia hopper package designed for specific task. the client however can still buy H100 to be paired with CPU of their choice though be it intel or AMD or IBM Power processor.

And its not X86. Fail.

lol why it was fail? the high performance segment are not dominated by x86. the fact that current fastest supercomputer in the world are based on ARM (Fujitsu Fugaku with their custom ARM A64FX) shows that ARM can also compete in the higher performance segment not just mobile SoC.

 

[raviadi1]

Looks like Nvidia is more worried about AMD than they are about Intel. Can tell because they didn't compare their chips to Xeon at all. Instead focusing on AMD's 2020 release of Epyc Rome. That doesn't exactly scream homerun, but impressive none the less.

Regardless, I am noticing Intel is becoming more and more an outsider in the server CPU space even by the likes of their own competition.

They actually worried about Cerebras which is their new direct competitor in AI Market because Cerebras claimed that bigger chip allows them to solve bandwidth limitation problem in AI Development, which is why Nvidia is trying so hard to buy Mellanox because of Infiniband and trying super hard for ARM deals to solve their front-end layer in AI Programming while mantaining that lightning-fast bandwidth in order to be scalable for cluster computing.

 

[dalek1234]

lol what is with both intel and nvidia cherry picking 2 year old AMD hardware to benchmark against the current offerings.

...

I know. Pretty desperate, isn't it? What's next, comparing your unreleased processor to your competitor's 10-year-old one? It's like listening to Putin

 

[jp7189]

because that's what they used on their DGX A100? the future DGX will use nvidia in house Grace ARM CPU.

No that's pure marketing fluff. The DGX H100 will be out this year before Grace will be ready. What CPU will be driving it? You can pretty much bet it will be Milan and they won't be highlighting that fact. Regardless, when upgrading their A100 system they had a choice to stay AMD (Milan) or go out on their own (Grace); that's the real comparison that matters.

 

[renz496]

No that's pure marketing fluff. The DGX H100 will be out this year before Grace will be ready. What CPU will be driving it? You can pretty much bet it will be Milan and they won't be highlighting that fact. Regardless, when upgrading their A100 system they had a choice to stay AMD (Milan) or go out on their own (Grace); that's the real comparison that matters.

nvidia just putting out some comparison to show that their Grace CPU is comparable to modern x86 CPU. they are not going to compare numbers to win the CPU performance crown like AMD and intel did. because those DGX machine majority of the workloads are going to be done on the GPU anyway. as i said earlier we will not going to see nvidia to sell their Grace CPU as a stand alone product in the server market. those Grace CPU are not going to compete directly with Xeons or Epycs. so comparing those Grace with the latest stuff from AMD and intel most likely not that important to those that want to buy those DGX system. also with H100 nvidia are adding DPX instruction support to the GPU which usually done on CPU or FPGA in the past. as mentioned by anandtech it is another type of workload that GPU are taking away from CPU.