AI isn't throttling HPC. It is HPC In recent discussions with industry vendor sales/marketing types, I've been hearing that HPC demand is falling off while AI system demand is continuing to increase. I've also seen articles implying that AI is somehow displacing HPC. Huh? Ok, this is The Reg, so I'd like to open this semi-rant by stating that the idea AI growth …
It would, I think, be more accurate to say that AI is replacing HPC. It uses the same kit, but it doesn't replicate the performance. So it's not itself HPC, but it's also not throttling HPC because HPC systems are still being created.
It's like using a Formula 1 car for your daily commute. It's completely unnecessary and a waste of the potential, but doing it doesn't mean that there's somehow fewer F1 cars.
It's them switching terms to simply continue the hype, which of course makes them MONEY. "It isn't HPC", they say, so they can hype you on buying an entirely new server farm to meet their new, glossy "AI" expectations. Not "add on", not "reconfigure", not "reallocate" - it's "all new" at the next refresh cycle with more outlay, because, of course, "AI" is all now a "necessity". It's desktop "CoPilot PC" hype on a 40kW scale.
In most of these examples it’s inference. Running the numbers to prove hypothesis and once you have your potential winners doing the science/engineering on them - whether BioChemistry for drug manufacturing and clinical trials, fluid dynamics for an F1 Front wing or NASA chewing through Petabytes of data.
Good old fashioned workloads, looking for stuff.
> Nvidia was talking about 1 megawatt per rack by the end of the decade. Sure hope the bubble bursts before that or here in the US we'll be dealing with electricity prices 4x higher than today and regular rolling blackouts like we're some kinda third world country.
With that amount of heat in a single any failure in the cooling system will have things melt down far faster than any protection system could reach
I was thinking how hard it would be to cable that but realised that if we do reach that point then power would be coming down busbars - just don't drop that spanner!
I'd like to see open foam or other 'techical' applications make any useful contributions with a 4 bit wide bus (even if there are x n of them in parallel). I know that is a stretch, but the cards for AI are dropping the wide / high floating point capacity in favor of the low crap, as has been reported here multiple times.
spot on really. Heavy RAM requirements? Check. Heavy storage requirements? Check. High speed interconnects? Check. And I'll note the newer supercomputers have (almost if not entirely) moved toward having GPUs available for compute. One can argue semantics but these AI clusters have very much in common with the traditional HPC builds.
Perhaps once the AI bubble bursts (I don't think AI will become irrelevant or anything, but really AI in your fridge and etc? Really...) some of these will be repurposed for high speed compute.
