NVIDIA Readies At Least 3 New Hopper AI GPUs to Bypass US Sanctions
The original article is as follows:
A report from Semianalysis suggests that NVIDIA is readying more GPUs to circumvent the recent AI ban on China. The US Government has tightened its stranglehold around almost every AI chip by introducing new parameters to identify outliers. These parameters take into account the performance density, TPP (Total Processing Performance) instead of just raw bidirectional bandwidth.
NVIDIA Prepping H20, L20 & L2 Lineup of AI GPUs
Semianalysis claims NVIDIA is working on 3 new AI GPUs based on the Hopper architecture. Exact specifications are not known yet, but we do have some numbers. The new GPUs include the H20 SXM, L20 PCIe, and L2 PCIe. The suffix identifies the interface of these GPUs.
Going over our limited details, the H20 SXM features 96GB of memory with a total bandwidth of 4 TB/s. It can output 296 TFLOPs of compute power, netting a performance density of 2.9.
Performance density is the TPP divided by the die size. The important part is that 5.92 is the absolute performance density limit and all chips above this threshold will be banned with immediate effect.
Likewise, the L20 features 48GB of memory with a bandwidth of 0.9 TBps. It has a performance density of 3.1 (for 239 TFLOPs), still in the safe zone. Last but not least, the NVIDIA L2 packs 24GB of memory and a relatively low bandwidth of just 0.3TB/s. Despite this, it has the highest performance density at 5.2, though it still dodges the ban by a hair’s breadth.
Interestingly, the same report suggests that one specific GPU is 20% faster than even the H100 in LLM inference. This is due to its similarities with the forthcoming H200.
Furthermore, one of the China specific GPUs is over 20% faster than the H100 in LLM inference, and is more similar to the new GPU that Nvidia is launching early next year than to the H100!
Semianalysis
So it seems that even such strict controls were not adequate to prevent the export of high-performance GPUs to China. There are simply too many factors to consider and it would be a hassle to take countermeasures against all of them.
Note that NVIDIA is yet to announce these GPUs formally. While this is just a leak, we expect a launch shortly.
Source: Semianalysis
