The Nvidia H100 is an incredibly powerful AIB with 14,592 CUDA cores and 80GB HBM3 memory, but it comes with a hefty price tag of $45,000. However, it lacks a display processor, making it unsuitable for gaming. Despite this limitation, the Chinese YouTube channel Geekerwan managed to use a secondary AIB to enable features like real-time ray tracing on the H100. Geekerwan obtained four smuggled units and thoroughly tested them, though the results were not particularly impressive. The H100 utilizes the GH100 processor, supporting various data formats for AI and HPC workloads. Nvidia does not prioritize optimizing Hopper drivers for gaming.
Even though the Nvidia H100 has an unbelievable 14,592 of SIMD (aka CUDA) cores, an outrageous amount of memory (80GB HBM3), and costs an obscene amount of money ($45,000), it is not designed for graphics or gaming because it doesn’t have a display processor. That didn’t stop the folks at the Chinese YouTube channel Geekerwan who used a secondary graphics AIB to trick the system into recognizing the setup and unlocking features like real-time ray tracing for the H100. Geekerwan actually got four smuggled units to try. And boy, did he put them through their paces. Geekerwan has over 329,000 subscribers.
The YouTuber host is understandably excited about having such expensive gear, and then goes into some of the significant differences between the H100 and the Nvidia A100. Later, he used an A100 in an RTX 4090 as a display processor.
The H100 AIB from Nvidia utilizes the GH100 processor, featuring an impressive 14,592 CUDA cores designed to handle diverse data formats crucial for AI and HPC workloads. These formats include FP64, TF32, FP32, FP16, INT8, and FP8. In contrast, Nvidia’s consumer GPUs, like the AD102, primarily excel in FP32 support. And even though the GH100 doesn’t offer a display processor, it has 24 raster operating units (ROPs) but lacks display engines or outputs. It’s important to note that Nvidia doesn’t prioritize optimizing Hopper drivers for gaming applications.
Geekerwan ran Time Spy on the H100 and cleverly used the RTX 4090 as the output display processor.
Geekerwan built, using a 3D printer, a special cooling system for the H100, and then hooked up all four AIBs and ran some AI and Blender tests. All of which were not very impressive but fun as hell.
Up until today, I used to say that when it comes to GPUs, you get what you pay for. Now I have to amend that to: you get what you pay for if you pick the right apps.