One would possibly assume {that a} GPU that prices over $40,000 goes to be the most effective graphics card for gaming, however the fact is much more advanced than that. The truth is, this Nvidia GPU can’t even sustain with built-in graphics options.
Now, earlier than you get too upset, it is best to know I’m referring to Nvidia’s H100, which homes the GH100 chip (Grace Hopper). It’s a strong information middle GPU made to deal with high-performance computing (HPC) duties — not energy PC video games. It doesn’t have any show outputs, and regardless of its intensive capabilities, it additionally doesn’t have any coolers. It’s because, once more, you’d discover this GPU in an information middle or server setting, the place it’d be cooled with highly effective exterior followers.
Whereas it “solely” has 14,592 CUDA cores (which is fewer than the RTX 4090), it additionally has an insane quantity of VRAM and an enormous bus. In whole, the GPU sports activities 80GB of HBM2e reminiscence, break up into 5 HBM stacks, every related to a 1024-bit bus. In contrast to Nvidia’s client GPUs, this card additionally nonetheless has NVLink, which implies it may be related to work seamlessly in multi-GPU programs.
The query stays: Why precisely is this sort of GPU fairly so dangerous at normal utilization and video games?
To display the case, YouTuber Gamerwan was given 4 of those H100 graphics playing cards to check, and determined to place one in an everyday Home windows system to take a look at its efficiency. This was a PCIe 5.0 mannequin, and it needed to be paired with an RTX 4090 as a result of lack of show outputs. Gamerwan additionally 3D-printed a custom-designed exterior cooler to maintain the GPU working easily.
It takes a bit of labor to even get the system to acknowledge the H100 as a correct GPU, however as soon as Gamerwan managed to beat the obstacles, he was additionally in a position to activate ray tracing assist. Nevertheless, as we later discover all through the testing, there’s not a lot assist for anything on a non-data middle platform.
In a regular 3DMark Time Spy take a look at, the GPU solely hit 2,681 factors. For comparability, the common rating for the RTX 4090 is 30,353 factors. This rating places the H100 someplace between the buyer GTX 1050 and GTX 1060. Extra importantly, it’s virtually the identical as AMD’s Radeon 680M, which is an built-in GPU.
The gaming assessments additionally went poorly, with the graphics card hitting a median of 8 frames per second (fps) in Purple Useless Redemption 2. The dearth of software program assist rears its ugly head right here — though the H100 can run at a most of 350 watts, the system can’t appear to push it previous 100W, leading to vastly throttled efficiency.
There are a number of completely different causes for this poor present of gaming powers. For one, althoughrthe H100 is an ultra-strong graphics card on paper, it’s a lot completely different on an architectural stage than the AD102 GPU that powers the RTX 4090. It solely has 24 raster working models (ROPs), which is a major downgrade from the 160 ROPs that the RTX 4090 has. As well as, solely 4 out of 112 texture processing clusters (TPC) can render graphics workloads.
Nvidia’s client GPUs obtain plenty of assist on the software program facet in an effort to run properly. This consists of drivers, but additionally system assist from builders — each in video games and in benchmark packages. There aren’t any drivers that optimize the efficiency of this GPU for gameplay, and the result’s, as you’ll be able to see, extraordinarily poor.
We’ve already seen the ability of drivers with Intel Arc, the place the {hardware} has remained the identical, however improved driver assist delivered efficiency beneficial properties that made Arc an appropriate alternative should you’re shopping for a finances GPU. With no Nvidia Sport Prepared drivers and a scarcity of entry to the remainder of Nvidia’s software program stack (together with the ever-impressive DLSS 3), the H100 is a $40,000 GPU that has no enterprise working any form of sport.
In essence, this can be a computing GPU, and never a graphics card in the identical method that almost all of us know them. It was made for all types of HPC duties, with a powerful deal with AI workloads. Nvidia maintains a powerful lead over AMD the place AI is anxious, and playing cards just like the H100 play a giant half in that.
Editors’ Suggestions
