By Greg Gao
(JW Insights) Nov 13 -- US chipmaker Nvidia reportedly will release new artificial intelligence chips with significantly reduced computing power aimed at the Chinese market to meet the region’s growing demand for AI technology while complying with US export controls, according to various media outlets.
These chips include HGX H20, L20 PCle, and L2 PCle. Industry insiders reveal that the latest trio of chips is a modified version of the H100 chip, with Nvidia expected to officially announce this development no earlier than November 16. Chinese tech companies are anticipated to receive the products in the following days.
Leaked information suggests these three new AI chips are “scaled-down” versions. The HGX H20, designed for AI model training, has limitations in bandwidth, computational speed, and other aspects. The overall computing power is theoretically reduced by around 80% compared to the Nvidia H100 GPU chip, meaning that the H20 offers only 20% of the comprehensive computing power of the H100.
Additionally, it includes HBM memory and NVLink interconnection modules to enhance computing power at a higher cost. Despite a decrease in price compared to the H100, the expected cost of the H20 is still projected to be higher than that of AI chips provided by domestic companies, like Huawei’s Ascend 910B, a newer version of the 910.
An industry expert explained the situation, likening it to widening a highway but leaving the toll booth entrance unchanged, limiting the flow. Technically, through hardware and software locks, precise control over the chip’s performance can be exerted without requiring large-scale production line changes. A task that previously took 20 days with the H100 might now take 100 days with the H20, the expert estimated.
Previous reports indicated that starting from November 17, 2023, Nvidia’s GeForce RTX 4090 graphics card would no longer be exported to China.
The US Department of Commerce, in October 2023, introduced export control measures specifying two parameters for AI chips: total computing power and performance density. These measures aim to prevent countries, including China, from accessing such high-performance chips.
Subsequently, the US government demanded that Nvidia, despite a one-month buffer period, immediately enforce this ban. The previously released Chinese-specific versions of Nvidia’s AI chips, the A800 and H800, designed to comply with previous export rules, are now also prohibited from export under the new regulations.
