Comments on: Inside The Massive GPU Buildout At Meta Platforms https://www.nextplatform.com/2024/03/13/inside-the-massive-gpu-buildout-at-meta-platforms/ In-depth coverage of high-end computing at large enterprises, supercomputing centers, hyperscale data centers, and public clouds. Wed, 17 Jul 2024 11:29:13 +0000 hourly 1 https://wordpress.org/?v=6.5.5 By: Timothy Prickett Morgan https://www.nextplatform.com/2024/03/13/inside-the-massive-gpu-buildout-at-meta-platforms/#comment-228027 Wed, 17 Jul 2024 11:29:13 +0000 https://www.nextplatform.com/?p=143809#comment-228027 In reply to King.

If you mean physical location, we don’t know.

]]> By: King https://www.nextplatform.com/2024/03/13/inside-the-massive-gpu-buildout-at-meta-platforms/#comment-228002 Wed, 17 Jul 2024 03:03:34 +0000 https://www.nextplatform.com/?p=143809#comment-228002 In reply to Timothy Prickett Morgan.

thanks, which one data center or two data centers house the two clusters used to train LLAMA3?

]]> By: Timothy Prickett Morgan https://www.nextplatform.com/2024/03/13/inside-the-massive-gpu-buildout-at-meta-platforms/#comment-227957 Tue, 16 Jul 2024 11:44:35 +0000 https://www.nextplatform.com/?p=143809#comment-227957 In reply to King.

Funny you should ask. Read this:

https://www.nextplatform.com/2023/09/26/meta-platforms-is-determined-to-make-ethernet-work-for-ai/

]]> By: King https://www.nextplatform.com/2024/03/13/inside-the-massive-gpu-buildout-at-meta-platforms/#comment-227919 Tue, 16 Jul 2024 06:03:15 +0000 https://www.nextplatform.com/?p=143809#comment-227919 Just out of curiosity, what are the data centers for the two 24,576 GPU data center class cluster deployments? What are the network requirements?

]]> By: Muzhi Li https://www.nextplatform.com/2024/03/13/inside-the-massive-gpu-buildout-at-meta-platforms/#comment-223555 Wed, 24 Apr 2024 13:28:41 +0000 https://www.nextplatform.com/?p=143809#comment-223555 Hi, I wonder if Meta is to buy 350,000 H100 chips or accumulate 350,000 by the end of the year.
As per his REEL, “we’re building massive compute infrastructure to support our future roadmap, including 350k H100s by the end of this year — and overall almost 600k H100s equivalents of compute if you include other GPUs”, it seems that 350k is the total H100 figure.
The beautiful table in this article may miss out MI300’s equivalent (2 MI300 = 1 H100?). Please check, thank you!

]]> By: Timothy Prickett Morgan https://www.nextplatform.com/2024/03/13/inside-the-massive-gpu-buildout-at-meta-platforms/#comment-221683 Thu, 14 Mar 2024 16:58:07 +0000 https://www.nextplatform.com/?p=143809#comment-221683 In reply to Slim Albert.

Yo, Slim. Don’t invoke “killer apps” please. . . .

]]> By: Slim Albert https://www.nextplatform.com/2024/03/13/inside-the-massive-gpu-buildout-at-meta-platforms/#comment-221664 Thu, 14 Mar 2024 12:33:07 +0000 https://www.nextplatform.com/?p=143809#comment-221664 Wow! Quite an impressive GenAI Infrastructure plan by Meta. The two 24,576-GPU datacenter-scale clusters each pack roughly twice the oomph of MS Eagle (#3 on Top500), suggesting about 1.3 EF/s each in FP64 (if Meta were so-inclined as to run one through HPL). In my mind, if “believable” AGI doesn’t emerge from such massive assemblies of artificial matrix-vector neurons, then we might be in for a recurrence of some AI-winter-like weather (global cooling of enthusiasm). Let’s hope it pans out (safely), with some successful and long-lasting killer-apps coming out of the effort!

]]>