Compared to the H100, how is the performance of the most powerful China-specific AI chip, the H20?

Almost certainly, all three AI chips are "castrated" or "shrunken" versions of H100, reflecting adjustments made to comply with export controls. Among them, the H20 AI chip has 96GB of HBM3 storage space, memory bandwidth of up to 4.0 Tb/s, are higher than the H100, but the comprehensive computing power is only 296 TFLOP, performance density of 2.9, far less than the H100.

In terms of specific performance metrics, the H20 AI chip is a kind of tweaked version of the H100. According to measurement organizations, the H20's combined arithmetic power is about 80% lower compared to the H100. This change reflects the chip's performance tweaks to accommodate requirements related to U.S. export control policies. Nonetheless, H20 still offers performance advantages in specific situations. For example, by reducing the number of chips required for inference from two to one, and if 8-bit quantization is then used, the LLAMA 70B model can be run efficiently on a single H20 instead of requiring two H100s. this demonstrates that the H20 can still provide effective performance in certain application scenarios.

From a traditional arithmetic perspective, H20 is a downgrade from H100, but in this aspect of LLM inference, H20 will actually be over 20% faster than H100, on the grounds that H20 is similar in some ways to H200, which is due for release in 2024. Note that the H200 is the successor to the H100, focusing on super-performance chips for complex AI and HPC workloads. As such, NVIDIA's H20 AI chip was launched in response to US export control policies towards China, and its performance has been reduced compared to the H100, but it still maintains a certain level of efficiency and utility in certain application scenarios.

NVIDIA and other big tech giants plan to unveil their latest developments at the Consumer Electronics Show (CES) in Las Vegas this week. The company is expected to showcase several of its latest GPUs at the event, including the RTX 4080 Super, 4070 Ti Super and 4070 Super. analysts generally expect that NVIDIA will tend to maintain its absolute leadership position in the global gaming hardware sector. $NVIDIA(NVDA.US)$ $Advanced Micro Devices(AMD.US)$