HGX H20, L20 and L2: NVIDIA to produce AI cards for China!


As you know, the Americans have imposed sanctions on China in the IT sector. The sanctions relate to graphics cards exceeding a certain power limit. The aim is to prevent the country from developing its military sector using American technologies. To achieve this, the American government is banning the export of cards exceeding 4800 TOPS of performance.

However, China represents a juicy market for NVIDIA, especially in the AI sector. Understandably, the chameleon doesn’t want to miss out on this market, and for this reason, the brand is planning new cards. Of course, these will comply with sanctions. These new cards are the HGX H20, L20 and L2!

HGX H20, L20 and L2: NVIDIA won’t give up on China just like that!

NVIDIA HGC H20 - L20 - L2As a reminder, the new directives prohibit NVIDIA from selling A100, H100, A800, H800 and RTX 4090 (whose price has skyrocketed) on Chinese soil. Nevertheless, in order to continue selling there, the company will soon be offering new solutions that comply with the sanctions. It has to be said that the Chinese are very keen on specialized AI cards.

NVIDIA HGX H20 - L20 - L2

On the program is the HGX L20, a model in HGX format based on Hopper architecture. This model offers a massive amount of memory: 96 GB in HBM3. What’s more, bandwidth is insane: 4Tb/s. In terms of computing performance, this model boasts 296 TFLOPS in INT8, 148 TFLOPS in BF16 via tensor cores, 44 TFLOPS in FP32 and 1 TFLOP in FP64. Finally, it features a PCIe 5.0 interface, as well as a 900 GB/s NVLINK link.

The L20 and L2 are PCIe cards equipped with 48 GB and 24 GB GDDR6. The respective memory bandwidths are 864 GB/s and 300 GB/s. Here, the GPU selected is an AD102, and the announced performance figures are as follows:

  • L20: 239 TFLOPS (FP8) – 119.5 TFLOPS in BF16 – 59.8 TFLOP in FP32
  • L2: 193 TFLOPS (FP8) – 96.5 TFLOPS in BF16 – 24.1 TFLOP in FP32

NVIDIA HGX H20 - L20 - L2

Ironically, according to Computer Base, the H20 is already faster than the H100 in certain areas. Clearly, the integration of new functions has helped. However, if we look at the figures, compared with the H100, these are seriously limited, particularly in terms of TFLOPS. Ironically, memory capacity is similar to, or in some cases even greater than, the sanctioned cards.