Comments on: A Quick Introduction to the NVIDIA GH200 aka Grace Hopper

By: eastcoastpete

eastcoastpete — Mon, 16 Sep 2024 04:23:51 +0000

Thanks, great overview. ChipsandCheese had an interesting look at the Grace CPU, and one of their Conclusions was that the Neoverse V2 -based Grace CPU showed some of the downsides of Nvidia using a not-customized ARM Neoverse V2 Design probably left performance on the table. https://chipsandcheese.com/2024/07/31/grace-hopper-nvidias-halfway-apu/

By: Good luck

Good luck — Mon, 09 Sep 2024 02:04:19 +0000

So if you’ve got the 480GB GH200 you don’t have insane memory bandwidth? The combined ranked DIMMs will hit these numbers on normal CPUs ez.

By: frogfreak

frogfreak — Sun, 08 Sep 2024 21:45:23 +0000

Best GH200 resource out there other than what NVIDIA’s own site’s got.

By: Hims

Hims — Sat, 07 Sep 2024 02:55:10 +0000

I’m shocked. We’ve got a few thousand H100s and I thought GH200 meant all were H200 type.

By: drikster80

drikster80 — Fri, 06 Sep 2024 23:46:13 +0000

Great write-up. I’ve been working with the GH200 for a couple of months and honestly really like it. The arm64 means a lot of newer apps needs to be built from source, but the performance is pretty good when clustering a few of them together. I definitively recommend using tcmalloc to utilize that CPUGPU memory transfer speed. When comparing 4xGH200/BF3 cluster to my 4xA100 watercooled system, it is more than twice the speed. Looking forward to seeing what the new GB200 brings with it.

By: francis

francis — Fri, 06 Sep 2024 23:05:35 +0000

I thought it was a product, not a family of products too.