Comments on: Top500 Supers: This Is Peak Nvidia For Accelerated Supercomputers

By: Bruce Considine

Bruce Considine — Wed, 15 May 2024 05:07:11 +0000

“…cheap and reliable fusion power…”
A fantasy that may never happen (ALWAYS 10-20 years away, the reports of nearing grid power are deceitful.) Meanwhile, we have cheap and reliable Molten Salt Reactors (fission) that have never been rolled out.

By: HuMo

HuMo — Tue, 14 May 2024 15:03:32 +0000

A true gastronomic HPC pleasure, for the finest of palates, to finally get a taste of the value-meal compute-combos as they freshly entered the menu at that Top500 hole-in-the-wall, with the GH200 grasshopper chef special, and the meditative instinct MI300A entree!

Well … that entree was really represented by identical triplets of appetizers, not enough to nourish a full-fledged swashbuckling buccaneer on its own, but an early deliveroo, with Grizzly Adams and the American Cro-Magnon of Tuolumne as conjoined computational acolytes. The El Capitan clearly decided to exhibit its stronger swagger in tip-toe motion this time, wearing a red hat (but no tutu), which is more pleasing than doing so in no-show motion as had happened before. Appetizers are meant to whet your appetite they say … well, am I thirsty for some serious rhumba now (and suplex)!

Meanwhile, the locust plague of the ARMageddon, the East Coast cicadas’ ARMpockyclypse and its 2024 double-brood of doom, them, they did show up with a gut-bomb army of swiss cheese graters, red-nosed rocket sleighs, and even telekinetic warrior-monks of the jedi persuasion — an impressive set of HPC kitchen staff and utensils, spanning the 4 to 270 range in cricket-pancake petaflips, per second! “Tastes great. Less filling”? No silly, tastes great, more filling! 8^b

By: Hubert

Hubert — Tue, 14 May 2024 11:09:42 +0000

Interesting angle and provocative suggestion! Seeing how shares of peak teraflops match those of total cores, one may infer that the best bang-for-the-buck will come from using an appropriate number of the least expensive cores (if any), irrespective of vendor (between AMD, Intel, and Nvidia) — effectively neutering potential high price premiums commandeered by the most publicized option. In the current context, this could indeed lead to peak Nvidia for this computational space (particularly when coupled with HPL-MxP leadership of Aurora and Frontier). Quite insightful!

One would probably also want to consider energetic efficiency of the cores, for TCO, as planned for upcoming articles in the introductory paragraphs.