In case you didn't know, AI and high performance computing is big business. There's a reason Nvidia has a market cap of over a trillion dollars. That reason is exploding demand for its enterprise products including the mighty H100 Hopper GPU. Yep, this monster processor, which can cost $30,000 or more, shares much of its DNA with humble GeForce gaming graphics cards.
As a fun experiment, YouTuber Geekerwan (via Tom's Hardware) took one of these monster cards for a spin through a couple of gaming benchmarks. The results were interesting, though ultimately good for a giggle if nothing else.
The H100 variant Geekerwan used was the PCIe version. It came equipped with 80GB of HBM2e memory, 14,592 CUDA cores and a 350W TDP. Compare that to an RTX 4090 with 24Gb of GDDR6X, 16,384 CUDA cores and a 450W TDP. The H100 shouldn't be a slouch right?
Actually it's very poor at gaming, with the card producing a 3DMark Time Spy graphics score of just 2,681. That's less than Radeon 680M integrated graphics. In Red Dead Redemption 2, the card couldn't even hit 30 FPS at 1080p.
In all seriousness, these results aren't surprising. While the H100 is an immensely powerful card, it's not designed for graphics applications. In fact, it doesn't even have display outputs. The system needed a secondary GPU to provide a display. It also lacks some other fixed hardware critical for gaming.
Then there's the fact the driver is completely unoptimized for gaming. During the gaming tests, the GPUs power consumption was sub 100W, indicating a major lack of utilization.
So, if you've got $30,000 to burn, buy a car. Or buy an RTX 4090 system and a car with the money you'll have left over. The mighty H100 card is not what you'd use to power the ultimate gaming rig.