Comments on: Money Changes Everything For SiPearl https://www.nextplatform.com/2023/04/05/money-changes-everything-for-sipearl/ In-depth coverage of high-end computing at large enterprises, supercomputing centers, hyperscale data centers, and public clouds. Wed, 19 Apr 2023 15:39:19 +0000 hourly 1 https://wordpress.org/?v=6.5.5 By: gc https://www.nextplatform.com/2023/04/05/money-changes-everything-for-sipearl/#comment-207187 Thu, 13 Apr 2023 13:07:50 +0000 https://www.nextplatform.com/?p=142209#comment-207187 Hot drink pointing at the hot center CPU stripe. Maybe the best cooling pipe will flow over that stripe?

]]> By: Hubert https://www.nextplatform.com/2023/04/05/money-changes-everything-for-sipearl/#comment-206883 Thu, 06 Apr 2023 15:01:35 +0000 https://www.nextplatform.com/?p=142209#comment-206883 In reply to Timothy Prickett Morgan.

Thanks again for the kind words! I’m looking forward to getting my mind blown clean-off by EU HPC designs, as it has been by Fugaku’s A64FX, and Frontier’s EPYC+MI250x, and will likely be as well (in June?) by GH100, SR/SP Max, and/or MI300A! If they put as much attention and creativity to this as they do to gastronomy, it should end-up in pure computational bliss!

]]> By: Timothy Prickett Morgan https://www.nextplatform.com/2023/04/05/money-changes-everything-for-sipearl/#comment-206857 Thu, 06 Apr 2023 00:36:41 +0000 https://www.nextplatform.com/?p=142209#comment-206857 In reply to Timothy Prickett Morgan.

And humorous Hu per usual.

]]> By: Timothy Prickett Morgan https://www.nextplatform.com/2023/04/05/money-changes-everything-for-sipearl/#comment-206856 Thu, 06 Apr 2023 00:36:06 +0000 https://www.nextplatform.com/?p=142209#comment-206856 In reply to Hubert.

Agreed on the L3 and cores, but there are still four cores without L3s at the corners then.

]]> By: Hubert https://www.nextplatform.com/2023/04/05/money-changes-everything-for-sipearl/#comment-206852 Wed, 05 Apr 2023 23:01:16 +0000 https://www.nextplatform.com/?p=142209#comment-206852 Great to see this update on the EU HPC ARM effort! If the L3 are the larger green squares, then I think each is paired with the core directly across the interconnect (white band) from it (rather than with the visually more-directly-accessible core — for some French reason). I wonder why they didn’t group the 24 top cores, and the 24 bottom cores, in the same way as the 24 central cores (with the shoelace interconnect) — if it’s good for the goose, it should be good for the “foie gras”! Yet, I think that I would suggest a quadtree organization of the cores (eg. 64 of them) and interconnect instead, to get O(log2(n)) intercore hopping, at low cost, Hanan Samet-style (former UMCP colleague) — a few of the FEA mesh generation, and broader matrix decomposition strategies covered in recent TNP articles were similarly based on quadtrees (2-D) and octrees (3-D). The pastry known as “mille-feuilles” could also, I think, inspire some tasty hyper-space connectivity (better than the accordéon) … hmmmm, yummy, … food for thought!

]]>