The US’s trade sanctions have restricted Nvidia from exporting its advanced GPUs like the A100 and H100 to China.
As a workaround, Nvidia developed the A800 and H800 GPUs specifically for Chinese clients, but these were also blocked by updated US controls last October, leading to the creation of the H20.
Specifically tailored for the Chinese market, this GPU is aimed at AI training and comes with a price tag of between $12,000 and $15,000, according to South China Morning Post which cited anonymous industry sources.
Falling short
Nvidia’s latest budget AI GPU has reportedly now landed in China, but it is facing stiff competition from its own previous models and Huawei’s secret weapon, the Ascend 910B. The H20’s computational capabilities are said to be similar to Huawei’s GPU, a chip developed by the Chinese tech giant which is currently under US trade sanctions. The Ascend 910B, while not officially confirmed by Huawei, is available through various sales channels in China.
The performance of Nvidia’s H20, however, appears to fall short of its predecessors. According to some sources, the H20’s capabilities lag significantly behind the A800 and A100, though full in-house testing is not yet complete.
In terms of price, a high-end server featuring eight Nvidia H20 cards is priced at about $195,800. In comparison, a Huawei server embedded with 910B chips sells for approximately $200,000 as of January this year.
While a few Chinese clients have reportedly received H20 samples for testing, and pre-orders have now opened, Nvidia’s GPU is expected to reach China in the second quarter, with shipments of products containing the H20 expected to start in July.
Nvidia’s GPUs are in high demand globally, with Chinese tech giants spending billions to acquire the GPUs to power new AI technologies. The company’s data center segment surpassed video gaming to become the top revenue contributor in its 2023 financial year, with a record $14.51 billion in revenue.
However, it remains to be seen how the H20 will be received by Chinese clients, given its constrained capabilities due to the new US rules.
