Chinese chipmaker, Tianshu Zhixin has successfully powered up the country’s first 7nm Data Center GPU aimed at neural network training. This is the first 7nm GPU fully designed and built domestically without any international assistance. There no confirmation on whether the GPU was fabbed by SMIC or some other Chinese foundry but considering the intricacies surrounding advanced processes, the former is the most likely responsible.
Shanghai Tianshu Zhixin Semiconductor Co., Ltd. on the 15th of January announced on its official website that the company’s flagship 7nm GPU (BI) for cloud computing and accelerating neural networks has been successfully powered on. According to STZS, this is the first fully self-developed “truly 7nm” GPGPU based on a home-designed GPU architecture. ()
Interestingly, according to the official website, the BI GPU leverages a 7nm process along with a 2.5D CoWoS packaging technology which as far as I know is TSMC’s 3D packaging IP. Overall, the HPC GPU packs a massive 24 billion transistors. In comparison, NVIDIA’s latest 7nm Ampere-based A100 accelerator packs 54 billion transistors while the AMD Arcturus GPU powering the Instinct MI100 includes 50 billion. Considering this is the first such product from the Chinese chipmaker, it’s nothing short of remarkable.