NVIDIA’s most expensive GPU has been tested in gaming
YouTuber Geekerwan has been given access to NVIDIA H100 Hopper graphics card.
In fact, it was not just one graphics card but four, each valued at 300,000 RMB (almost 42,000 USD). However, the H100 is not precisely a graphics card by itself, but an GPGPU (General Purpose) GPU or AI-accelerator for advanced data-center workloads. Yet, nothing should prevent anyone from using a PCI Express version on a normal desktop PC, which is what happened in this case.
NVIDIA H100 features a cut-down GH100 GPU with 14592 CUDA cores. The card makes use of 80GB HBM3 of capacity attached to a 5120bit memory bus (five HBM stacks each attached to a 1024 bit bus). This means that the maximum bandwidth of this GPU is 2TB/s. The PCIe version is actually one of the first NVIDIA GPUs to utilize the PCI Express 5.0 interface, not available on gaming series.
More importantly, the H100 PCIe does not have a fan. It’s a data-center GPU to be installed in rack servers, so external cooling is required. The YouTuber simply added a custom-made blower fan at one end, which was enough to cover the 350W TDP for this model (bear in mind the SXM version is rated at 700W).
Running H100 on such a system requires some work. The primary problem is the lack of any display output. This means that a secondary graphics card is required. Furthermore, data-center (former Tesla) cards do not work like GRID series (for cloud game streaming), but there is a way to trick the system into recognizing these cards a such. This is precisely what was done to enable the GPU, which also unlocks ray tracing support.
The H100 can run games, but performance leaves a lot to be desired. The system struggles to put H100 into high-power mode, so one can observe sub-100W power consumption and performance matching entry-level gaming GPUs like GTX 16 series or even Radeon 610M integrated graphics.
The H100 has much fewer raster operating units (ROP) than graphics cards like RTX 4090 (160 vs. 24), which is a bottleneck for such workloads. Furthermore, only 4 out of 112 TPC (Texture Processing Cluster) are capable of rending graphics workloads and there are no game-optimized drivers for the data-center series.
The H100 is not a gaming card and NVIDIA made sure of it. This video teaches about the differences between gaming and data-center series. It also explains why one shouldn’t use those expensive graphics cards to play games. The video has subtitles and viewers can select auto translation to desired language. It is definitely worth a look.
Source via @I_Leak_VN: