Jump to content

Talk:Hopper (microarchitecture)

Page contents not supported in other languages.
From Wikipedia, the free encyclopedia

Hopper or Grace?

[edit]

It seems NVidia now also calls it the Grace CPU: https://www.nvidia.com/en-us/data-center/grace-cpu/. So what is it? Jalwikip (talk) 14:08, 18 January 2023 (UTC)[reply]

The Grace CPU is something completely different. elijahpepe@wikipedia (he/him) 02:50, 30 May 2023 (UTC)[reply]

70 pound / 30 kg a H100? Really?

[edit]

Are we speaking about the chip itself? with the heat-sink? or the Node? Looks too much, not that I've hold one on hand, but for one chip that's too much. Especially that when I quote Jensen Huang in the article : "People think that Nvidia GPUs is like a chip. But the Nvidia Hopper GPU is 35,000 parts. It weighs 70 pounds" Looks like he is speaking about the chip. Since the news site not specialized/generalist, an error / misunderstanding might have slipped and could have been 17 pounds / 7kg instead of 70 pound / 30 kg, which look more suitable. Nexoft (talk) 18:10, 27 February 2024 (UTC)[reply]

This quote was from the earnings call. I believe it to be the correct quote. ShadyCrack (talk) 01:07, 13 March 2024 (UTC)[reply]
The quote is essentially accurate, but it's taken out of context, and even without that context, it does not support the article's assertion that "Hopper contains 35,000 parts and weighs 70 lb (32 kg)," which is simply not true.
First, Hopper is a microarchitecture, and microarchitectures have no inherent weight, although their specific implementations may. (This is analogous to the way that, say, chess has no inherent weight, although chessboards and printed books about chess do.)
Secondly, the quote came from the Nvidia CEO's extemporaneous response to an investment analyst's question about supply chain constraints. In his response, the CEO illustrates the complexity of Nvidia's supply chain by describing how many components go a single Nvidia product. He begins this segment by talking about board components, interconnection, and cabling, then says, "the NVIDIA Hopper GPU has 35,000 parts [and] weighs seventy pounds," then continues talking about cabling. Despite his choice of words, the full context of his response makes it clear that he's talking about a system built around one or more GPUs implementing the Hopper microarchitecture, not a single GPU on its own.
Finally, Nvidia sells multiple products containing Hopper-implementing GPUs that weigh drastically less than 70 lbs per GPU. The Nvidia H100 PCIe GPU contains one Hopper-implementing GPU and weighs 2.6 lbs. The Nvidia DGX H100 system contains eight Hopper-implementing GPUs and weighs 287.6 lbs, less than 36 lbs per GPU.
Trogon-nightjar (talk) 23:00, 13 April 2024 (UTC)[reply]