The Tianshu Zhixin company announced China’s first 7nm data center GPU, which will rival the options of Nvidia and AMD.
Tianshu Zhixin announced its Big Island GPGPU (BI) , which is touted as the first GPGPU in the Asian country that is designed for AI and HPC applications, education, medicine and security.
This GPGPU, also known as ‘BI’, has 24 billion transistors built under a 7nm node with 2.5D CoWoS packaging. Although they do not give details about who has manufactured these wafers, it is most likely TSMC, since Tianshu Zhixin do not have the capacity to manufacture them themselves.
This has been a two-year journey for the Chinese company, which started the development of BI chips in 2018. The final development of all the features occurred in May 2020, so mass production should have already occurred, thinking of meeting their goal of launching them for sale later this year.
Tianshu Zhixin promises FP16 performance of up to 147 TFLOPS. This would be above the Nvidia A100’s performance of 77.97 TFLOPS, but below the Radeon Instinct MI100 of 184.6 TFLOPS, although it lacks the Tensor cores that the A100 does.
Lastly, the GPGPU BI supports FP32, FP16, BF16, INT32, INT16, and INT8 floating point formats, just to mention a few of the most popular.
In this way, China is increasingly independent of Western technology, a trend that has been increasing in recent years.