Comments on: Nvidia Enters The Arms Race With Homegrown “Grace” CPUs https://www.nextplatform.com/2021/04/12/nvidia-enters-the-arms-race-with-homegrown-grace-cpus/ In-depth coverage of high-end computing at large enterprises, supercomputing centers, hyperscale data centers, and public clouds. Tue, 31 May 2022 16:58:47 +0000 hourly 1 https://wordpress.org/?v=6.5.5 By: Timothy Prickett Morgan https://www.nextplatform.com/2021/04/12/nvidia-enters-the-arms-race-with-homegrown-grace-cpus/#comment-161818 Wed, 14 Apr 2021 01:42:41 +0000 http://www.nextplatform.com/?p=138219#comment-161818 In reply to ChrisGX.

What he said was actually weirder than I thought. I am getting further clarification.

]]> By: Timothy Prickett Morgan https://www.nextplatform.com/2021/04/12/nvidia-enters-the-arms-race-with-homegrown-grace-cpus/#comment-161817 Wed, 14 Apr 2021 01:41:34 +0000 http://www.nextplatform.com/?p=138219#comment-161817 In reply to ChrisGX.

I meant a pair of them. Test ranges, which I looked up, were around 350. For two, not one. Apologies.

]]> By: ChrisGX https://www.nextplatform.com/2021/04/12/nvidia-enters-the-arms-race-with-homegrown-grace-cpus/#comment-161814 Tue, 13 Apr 2021 23:43:22 +0000 http://www.nextplatform.com/?p=138219#comment-161814 >> 300 on the SPECrate2017_int_base test, which is on par with a 28-core “Skylake” and “Cascade Lake” Xeon SP processor from Intel.

I doubt the Cascade Lake Xeon SP performs that well on the SPECrate2017_int_base test. Please check the benchmark information at the URL below. The Xeon Platinum 8280 is one of the best performing CPUs of the Cascade Lake Xeon SP range and I think its score is more like 100, not 300.

https://www.anandtech.com/show/16594/intel-3rd-gen-xeon-scalable-review/6

]]> By: ChrisGX https://www.nextplatform.com/2021/04/12/nvidia-enters-the-arms-race-with-homegrown-grace-cpus/#comment-161813 Tue, 13 Apr 2021 22:44:12 +0000 http://www.nextplatform.com/?p=138219#comment-161813 >> Huang said that four Grace CPUs would deliver over 2,400 on the SPECrate2017_int_base test over four nodes, which means it is 600 per CPU.

That’s not what I heard in the Keynote. He said 2,400 SPECrate2017_int_base would be delivered over 8 nodes which tallies with the 300 SPECrate2017_int_base claim made about the throughput of a single Grace CPU.

]]> By: A casual observer https://www.nextplatform.com/2021/04/12/nvidia-enters-the-arms-race-with-homegrown-grace-cpus/#comment-161810 Tue, 13 Apr 2021 18:20:37 +0000 http://www.nextplatform.com/?p=138219#comment-161810 The Denver-lineage of cores were still used in various Tegras until a few years back. Tegra X1, X2, used Denver follow-ons in mobile and Tegra Xavier used them in automotive.

]]> By: Timothy Prickett Morgan https://www.nextplatform.com/2021/04/12/nvidia-enters-the-arms-race-with-homegrown-grace-cpus/#comment-161795 Tue, 13 Apr 2021 01:44:40 +0000 http://www.nextplatform.com/?p=138219#comment-161795 In reply to Matt.

Absolutely right. In my rush, I forgot about the TensorCores. . . .

]]> By: Matt https://www.nextplatform.com/2021/04/12/nvidia-enters-the-arms-race-with-homegrown-grace-cpus/#comment-161791 Mon, 12 Apr 2021 21:06:58 +0000 http://www.nextplatform.com/?p=138219#comment-161791 I’m guessing the 20 exaflops of AI performance for the Alps machine means fp16 tensor core performance, not general purpose fp16 performance. So if the next generation GPU architecture is similar to Ampere in terms of its tensor core acceleration factors, the linpack fp64 performance will be much lower, under an exaflops. NVIDIA calls Selene a 2.8 AI-exaflops machine, and its linpack RMax is 63.5 petaflops. If we multiply that by 7 times we arrive at a 444.5 petaflops machine.

Regarding the graphic NVIDIA provided that lists the memory bandwidths, it lists an aggregated GPU bandwidth of 8000 GB/s for 4 GPUs, or 2 TB/s per GPU, which is the same bandwidth as the 80GB version of the A100. I think HBM3 should potentially be along by the time of Ampere-next, but even if not, there should be a memory bandwidth boost for the chip using HBM2e. Presumably the listed aggregate GPU bandwidth is not really what they expect to see in the product.

]]> By: EC https://www.nextplatform.com/2021/04/12/nvidia-enters-the-arms-race-with-homegrown-grace-cpus/#comment-161786 Mon, 12 Apr 2021 16:44:51 +0000 http://www.nextplatform.com/?p=138219#comment-161786 the real time stock charts of NVDA AMD and INTC are fore telling

]]>