Home » HGX H20, L20 and L2: NVIDIA to provide AI playing cards for China!

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

by Genzo
0 comment
NVIDIA HGC H20 - L20 - L2


As you realize, the Americans have imposed sanctions on China within the IT sector. The sanctions relate to graphics playing cards exceeding a sure energy restrict. The purpose is to stop the nation from creating its navy sector utilizing American applied sciences. To obtain this, the American authorities is banning the export of playing cards exceeding 4800 TOPS of efficiency.

However, China represents a juicy marketplace for NVIDIA, particularly within the AI sector. Understandably, the chameleon doesn’t need to miss out on this market, and for that reason, the model is planning new playing cards. Of course, these will adjust to sanctions. These new playing cards are the HGX H20, L20 and L2!

HGX H20, L20 and L2: NVIDIA gained’t hand over on China similar to that!

NVIDIA HGC H20 - L20 - L2As a reminder, the brand new directives prohibit NVIDIA from promoting A100, H100, A800, H800 and RTX 4090 (whose value has skyrocketed) on Chinese soil. Nevertheless, to be able to proceed promoting there, the corporate will quickly offer new options that adjust to the sanctions. It needs to be stated that the Chinese are very eager on specialised AI playing cards.

NVIDIA HGX H20 - L20 - L2

On this system is the HGX L20, a mannequin in HGX format primarily based on Hopper structure. This mannequin gives a large quantity of reminiscence: 96 GB in HBM3. What’s extra, bandwidth is insane: 4Tb/s. In phrases of computing efficiency, this mannequin boasts 296 TFLOPS in INT8, 148 TFLOPS in BF16 by way of tensor cores, 44 TFLOPS in FP32 and 1 TFLOP in FP64. Finally, it incorporates a PCIe 5.0 interface, in addition to a 900 GB/s NVLINK hyperlink.

The L20 and L2 are PCIe playing cards outfitted with 48 GB and 24 GB GDDR6. The respective reminiscence bandwidths are 864 GB/s and 300 GB/s. Here, the GPU chosen is an AD102, and the introduced efficiency 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, in line with Computer Base, the H20 is already quicker than the H100 in sure areas. Clearly, the mixing of latest features has helped. However, if we have a look at the figures, in contrast with the H100, these are severely restricted, notably by way of TFLOPS. Ironically, reminiscence capability is much like, or in some circumstances even higher than, the sanctioned playing cards.

You may also like

Leave a Comment