According to information on T-Head’s website, the Zhenwu 810E was a “fully self-developed” parallel processing unit (PPU), an application-specific integrated circuit designed for both AI training and inference. Alibaba owns the SCMP.
Developed by T-Head to handle the massive data throughput required for generative AI systems, the Zhenwu 810E supports 96 gigabytes of High Bandwidth Memory 2 Enhanced – a 3D-stacked synchronous dynamic random-access memory standard designed for AI, machine learning and high-performance computing.
A single card with the T-Head PPU handles a bandwidth of 700 gigabytes per second, which made it on par with Nvidia’s H20 – a graphics processing unit (GPU) tailored by Nvidia specifically for China to comply with US export restrictions .
China last year summoned Nvidia over alleged “back door safety risks” in the H20, which signalled a de facto ban on its sales to local firms.
T-Head’s launch of the Zhenwu 810E marked the latest initiative by China’s Big Tech firms, including Huawei Technologies and Baidu , to develop a domestic alternative to Nvidia’s GPUs and bolster the country’s technology self-reliance efforts.