Comments on: A GPU Upgrade For “Leonardo” Supercomputer But Not A Budget Upgrade https://www.nextplatform.com/2024/09/23/a-gpu-upgrade-for-leonardo-supercomputer-but-not-a-budget-upgrade/ In-depth coverage of high-end computing at large enterprises, supercomputing centers, hyperscale data centers, and public clouds. Wed, 02 Oct 2024 20:18:56 +0000 hourly 1 https://wordpress.org/?v=6.7.1 By: Slim Albert https://www.nextplatform.com/2024/09/23/a-gpu-upgrade-for-leonardo-supercomputer-but-not-a-budget-upgrade/#comment-235554 Wed, 25 Sep 2024 00:19:37 +0000 https://www.nextplatform.com/?p=144736#comment-235554 Interesting analysis! It seems that based on price, the A100-based system (similar tech to that already in Leonardo) would be just about the only choice (at 165 nodes), which is unfortunate given the improvements brought by Hopper and Blackwell. There’s also a power consumption limit in the “bidding requirements” of 1.2 MW IT in operation (and 1.6 MW during acceptance — sections 3.3.2-1, 3.3.2-2) which could limit options (again at 165 nodes). If the DGX H100/200 is 10.2 kW, then 165 nodes of that will work only during acceptance (plus be too expensive).

I’m with you then that it would be more sensible to specify a target “specific peak performance” than a set number of nodes. Just 70 nodes of Blackwells for example would probably give them better oomph at lower power than 165 nodes of A100s, at the same cost. It would also leave some room available in the floor plan where the heat generated could be used to cook local pasta, some ragù bolognese for example (eh-eh-eh!).

]]> By: Mehdi Zoghlami https://www.nextplatform.com/2024/09/23/a-gpu-upgrade-for-leonardo-supercomputer-but-not-a-budget-upgrade/#comment-235403 Mon, 23 Sep 2024 18:23:45 +0000 https://www.nextplatform.com/?p=144736#comment-235403 I hope they don’t add a Caravaggio to the mix.

]]>