NVIDIA’s AI-Conquering H100 Datacenter GPU Is Powerful, But It Probably Couldn’t Run Crysis
NVIDIA’s AI-Conquering H100 Datacenter GPU Is Powerful, But It Probably Couldn’t Run Crysis
Getting one of the coveted H100 GPUs outside of datacenter environment is no small feat in itself, but Chinese YouTuber Geekerwan (极客湾) has got ahold of not one, but four of the NVIDIA H100 GPGPUs (general-purpose graphics processing units) to take it for a spin.
One interesting fact to note, NVIDIA has stopped selling H100 GPUs recently due to US’s export sanctions towards China, and to comply with regulations, Chinese datacenter customers are now only allowed to access the “nerfed” version of the GPU named H800, with less compute power available on tap. That said, some companies has already obtained H100 before the sanction came into effect, making the H100 potentially an even rarer piece in the Chinese tech circle.
The GPU costs RMB300,000 each, or nearly RM200,000 for each card. Four of them and you’re looking at RM800,000 worth of hardware that occupies four PCIe slots, with each card drawing a relatively paltry 350W of power – half of the SXM variants of the same model. What happens when you put world’s most powerful GPU in games, then?
Turns out, not great. As mentioned earlier, the H100 is not GPU in the most traditional sense – but rather, it’s a compute-focused, general purpose card that excels in highly parallelized computing workloads, but the essential components for games to utilize its performance, such as ROPs and TMUs, are greatly cut down (it didn’t even contain video outputs for one). It’s also worth mentioning that H100 runs on an entirely different GPU architecture altogether, codenamed “Hopper”, hence not of the same lineage as consumer-focused Ada Lovelace-based RTX 40 GPUs.
Though, more surprisingly, gaming performance is practically abysmal considering the sheer raw power it contains. On 3DMark Time Spy it only matched the onboard graphics, specifically the AMD Radeon 680M that sits inside Ryzen mobile chips; and actual gaming workload sees it matching the mere Radeon 610M in performance. If you want to see the GPU in action, you can watch the video below (Chinese subtitles available, you can use YouTube’s auto-translate feature should you need to):
Source: Videocardz
Pokdepinion: Of course, this is just a fun experiment – but it’s always nice to see server-grade hardware thrown into “unconventional” use cases such as this one.
