Tech war: how Chinese scientists rigged a low-cost AI computer chip to power a hypersonic weapon | South China Morning Post
Tests have suggested that this particular module can process computational fluid dynamics models with unprecedented efficiency, meaning calculations that previously took seconds to complete could now be done in just 25 milliseconds – four times faster than the blink of an eye.
The module’s response speed made it ideal for “real-time optimisation of the fuel supply system, fault diagnosis, and fault-tolerant control in scramjet engines”, according to a joint project team from Beijing Power Machinery Research Institute and Dalian University of Technology. Their peer-reviewed paper was published on March 13 in the Chinese academic journal Propulsion Technology.
Headquartered in the United States, Nvidia is the world’s largest supplier of AI chips. Nvidia began selling the TX2i for industrial applications about six years ago. The module’s peak single-precision performance is 1.26 TFlops, about one-fiftieth of the capability of the company’s most powerful AI chip, the H100.
The H100, however, costs tens of thousands of US dollars and is in short supply. The TX2i, on the other hand, can be obtained for a few hundred dollars, is not subject to US export controls and is widely available online.
