YouTuber Geekerwan has recently obtained four NVIDIA H100 Hopper graphics cards, each valued at around 300,000 RMB (approximately 42,000 USD). Although the H100 is not specifically a graphics card, it functions as a GPGPU (General Purpose Graphics Processing Unit) or AI-accelerator designed for advanced data-center tasks. Interestingly, Geekerwan was able to use a PCI Express version of the H100 on a regular desktop PC, which is how they utilized it in this situation.
The NVIDIA H100 incorporates a reduced version of the GH100 GPU, boasting 14,592 CUDA cores. The graphics card is equipped with 80GB of HBM3 memory capacity, connected to a 5120-bit memory bus. This configuration includes five HBM stacks, each connected to a 1024-bit bus, resulting in a maximum bandwidth of 2TB/s for this GPU. Notably, the PCIe version of the H100 is among the early NVIDIA GPUs to utilize the PCI Express 5.0 interface, which is not typically available in gaming series cards.
Significantly, the H100 PCIe variant lacks an integrated fan as it is intended for installation in rack servers within data centers, where external cooling is necessary. The YouTuber addressed this by attaching a custom-made blower fan at one end, effectively meeting the cooling requirements for this model with a 350W TDP (thermal design power). It’s worth noting that the SXM version of the H100 has a higher power rating of 700W.
Operating the H100 on such a setup demands some effort. The main challenge is the absence of any display output, necessitating the use of a secondary graphics card. Additionally, it’s important to note that data-center cards (previously known as Tesla) function differently from GRID series cards designed for cloud game streaming. However, there is a method to manipulate the system into recognizing these cards as such. This was precisely implemented to activate the GPU, thereby enabling ray tracing support as well.
The gaming performance of the H100 falls short of expectations. The system encounters difficulties when attempting to activate the H100 in high-power mode, resulting in power consumption below 100W and performance comparable to entry-level gaming GPUs like the GTX 16 series or even Radeon 610M integrated graphics.
A notable limitation of the H100 is its significantly fewer raster operating units (ROP) compared to graphics cards like the RTX 4090 (160 vs. 24), which becomes a bottleneck for such workloads. Additionally, only 4 out of 112 TPC (Texture Processing Cluster) are capable of rendering graphics workloads, and there are no game-optimized drivers available for the data-center series.
The H100 is intentionally designed by NVIDIA as a non-gaming card, and this distinction is emphasized. The video below provides valuable insights into the disparities between gaming and data-center series cards. It also highlights why it is not advisable to utilize these costly graphics cards for gaming purposes. The video offers subtitles, and viewers can select automatic translation to their preferred language. It is definitely recommended to take a look at this informative resource.
Images Credit: Geekerwan