On May 26th, it was reported that NVIDIA has reduced the prices of its special AI chip H20 series in response to weak demand in the Chinese market.
According to three sources in the supply chain, Chinese server distributors are currently selling the H20 chips for around 100,000 RMB per set, with servers containing eight chips priced between 1.1 to 1.3 million RMB.
In some cases, these prices are more than 10% lower than Huawei's Ascend 910B chip.
Back in February, when NVIDIA's H20 was available for pre-order in China, the channel pricing ranged from $12,000 to $15,000, with some channel partners quoting prices as high as around 110,000 RMB.
NVIDIA's H20 chip, designed for the Chinese market, has reduced performance due to export control restrictions from the United States.
The H20 is a cut-down version of NVIDIA's H100 GPU, featuring 96GB of memory, a speed of up to 4.0Tb/s, computing power of 296TFLOPs, and a performance density of only 2.9.
In comparison, the H100 boasts 80GB of HBM3 memory, a memory bandwidth of 3.4Tb/s, computing power of 1979TFLOPs, and a performance density of 19.4. This means that the AI computing power of the H20 is less than 15% of the H100, with some aspects even falling short of the Ascend 910B.
Market analysts point out that Huawei is set to significantly increase shipments of the Ascend 910B chip this year, with the chip outperforming the H20 in certain key metrics.
Founder of research firm SemiAnalysis, Dylan Patel, predicts that by the second half of 2024, nearly one million H20 chips will be shipped to China. Nvidia must compete with Huawei on pricing to secure enough orders in the Chinese market to absorb this volume of chips.
