> Is it exclusively used for LLM training?

While LLMs is what cool kids all talk about nowadays, there's way, way, way more to the machine learning than them. I'd say that minority of GPUs are used for LLMs in the ML space. There's likely way more used them where they directly bring billions of dollars to the companies (like all the rankers across big tech, for content, ads, search, etc).

Are they price-competitive for those applications though? My understanding is that the main focus is Memory/GPU-interconnect for large models.

Are recommendation/ranking models large enough to take advantage of this? I don't think these cards are generally competitive in throughput/$.

It's not really clear what you mean, maybe there's gaps in my ML knowledge, but generally the answer depends on "what throughput do you need for your use case?"

Generally I agree that if you only need to deliver mail once daily on a mile long route, a car that takes 24 hours to travel a mile is fine.

Lots of ML applications don't generally care about latency, and throughput only matters per dollar (as you can just scale horizontally).

To my knowledge, most inferencing (at least for simpler models) happens on cheaper, slower GPUs that have better throughput/dollar.

Don’t take this the wrong way but have you been drinking?

? Have you? (Note I'm not op)

Do you have any idea of the scale of ranking infrastructure at big tech and how it’s been all ML for about a decade across the industry?

Sorry, it was a rude comment to make based on some strange grammar I noticed. I compared with some of your prior comments and noticed you didn’t have any such issues before. I wasn’t being facetious or judgemental - genuinely curious. There’s not a non rude way to ask that, should have kept it to myself.

Supercomputers also use them, like the 4th, 5th, 6th, 8th, and 9th largest ones.

The list of supercomputers using H100’s continues past #9

> TDP 450W - 1000W

Also useful as a space heater.

Literally. 1000w heaters going for $20 on amazon. Good savings!

That's practically the TDP for an entire x1 chassis though. Check out the TDP for a H100 DGX Pod. The performance per watt of GH200 *should* be a massive improvement over x86 with Hopper, especially for clusters

Not only training. Inference as well.

I feel like this is more and more in the majority too. So much so that AMD is talking about their chips for inference as much as training [0].

0. https://www.theverge.com/23894647/amd-ceo-lisa-su-ai-chips-n...

Although it is not as hyped, HPC is much wider than just machine learning. H100 are incredibly useful in numerical simulation and scientific computing.

Already answered below, but the bulk of the market for H100’s appears to be AI. HPC was historically the largest market for this class of accelerator, but it appears to have shifted to hyper scale cloud computing centers.

HPC’s are used to distribute the computation of differential equations using numerical methods and AI appears to be making inroads here as well. I would be interested in others’ opinion on whether it’s a settled question on whether or not AI will be able to produce equivalent solutions.

It's an iterative thing! Neural Network based surrogate models are becoming popular, but you need the compute to solve the PDEs to train the AI and compute the loss in the first place ;)