NVIDIA’s China-Compliant H20 GPU Has 41% Fewer Cores & 28% Lower Performance Versus Top Hopper H100 Config

NVIDIA's H20 Hopper GPU for China has been tested and features 41% fewer cores than the top H100 configurations.

Last year, it was reported that NVIDIA was preparing export-compliant GPUs for China to meet the surging AI demand in the region. The company finally started taking orders for these chips back in February and the prices were disclosed to be around $12,000 per piece for this cut-down variant. It is now revealed just how much of a cut-down chip and what kind of performance it offers though in a single synthetic test.

Officially, the NVIDIA Hopper H20 GPU was designed for the HGX platform with 96 GB of HBM3 memory, offering 4.0 TB/s of bandwidth versus the 8 TB/s bandwidth offered by the top solution. This chip featured 296 TFLOPS of INT8, 148 TFLOPS of FP16, 74 TFLOPS of TF32, 44 TFLOPS of FP32, and just 1 TFLOPS of FP64 performance. Other specs include support for up to 7 multi-instance GPUs, 60 MB L2 cache per chip, 8-Way HGX configurations, NVLINK bandwidth of 900 GB/s, and a 400W TDP.

NVIDIA never disclosed the GPU core count but based on a recent Geekbench 6 entry, it looks like we finally know what the chip has to offer. According to the data, NVIDIA's Hopper H20 AI GPU for China features just 78 SMs out of the total 144 SMs for the H100 GPU. The fastest H100 GPUs are equipped with 114 and 132 SMs so that's already a 41% drop in core count versus the top dies that are accessible to everyone else besides China. Furthermore, the GPU comes with 96 GB of HBM3 memory which is an upgrade from the original 80 GB HBM3 models, the newer H100 configs feature 144 GB of HBM3e memory.

As for performance, well the NVIDIA H20 Hopper GPU scores 248,992 points in the Geekbench 6 OpenCL test and while the GPU is made first for AI applications, it kind of gives us a hint at where the performance falls. The H100 80 GB configurations offer around 280K points in the same test while the 132 SM variants offer up to 350K points. The H100 144 GB configs should offer even higher performance so that should give you an idea of just how much of a downgrade NVIDIA had to come up with just to meet US regulations.

But despite the cut-down nature of the NVIDIA H20 Hopper GPU, it looks like customers in China are lining up to get these solutions. It was recently reported that NVIDIA is eyeing up to $12 Billion in terms of revenue despite restrictions thanks to its export-compliant chips such as the H20 and well the reason is the inability of domestic AI chipmakers, such as Huawei, to keep up with the growing AI demand.

It is said that the domestic AI demand within China is huge and Huawei can't fulfill orders for every customer which has made tech giants look for alternatives like the ones offered by NVIDIA. It is also said that despite being a significantly cut-down variant, NVIDIA's H20 AI GPUs still offer better performance capabilities and software support than the local AI tech competitors so it's a win-win for the green team.

News Source: Benchleaks