Detailed pictures of China's first competitive 7 nm compute GPU emerge, mass production starting soon
A few months ago, Tianshu Zhixin was announcing its progress on the first commercial GPGPU produced in China, noting that the global semiconductor shortages forced a rescheduled launch for the Big Island GPGPU that was supposed to be available in 2H 2020. These problems were sorted out in the meantime and the Shanghai-based company was able to soft launch its compute GPU on March 31. At the soft launch party, Tianshu Zhixin also showed new product pictures and provided additional performance details.
Despite its late launch, Tianshu Zhixin claims that the Big Island is still 1-2 years ahead of other similar Chinese endeavors, and this represents an important step for China as it continues to reduce its reliance on U.S. technology. China is already making great progress with its domestic fabrication nodes, yet the 24-billion transistor GPGPU itself is not produced by any Chinese fab. Tianshu Zhixin revealed that the chips are made using TSMC’s 7 nm nodes with 2.5D CoWoS packaging. The Big Island chip will integrate TSMC’s 65 nm interposers and the PCB will be fitted with 32 GB of HBM2 VRAM capable of 1.2 TB/s bandwidth. Furthermore, the board comes with a single 8-pin connector, supports 16x PCIe 4.0 connections and gets a 300 W TGP rating.
As far as performance goes, the Chinese GPGPU seems to be trading blows with AMD’s Instinct MI100 compute accelerator. We already knew that the Big Island is a bit slower in FP16 loads (147 TFLOPs vs 184.6 TFLOPs), but it is actually faster in FP32 loads (37 TFLOPs vs 23.1 TFLOPs). When it comes to integer calculations, the new GPGPU does not disappoint with 37 TOPS for Int32, 147 TOPS for Int16 and 295 TOPS for Int8.
Tianshu Zhixin also presented server layouts with 12 GPGPU clusters. Mass production is expected to begin shortly, but pricing info has not yet been revealed, although it was previously mentioned that the Big Island should have reduced costs compared to AMD’s MI100 or Nvidia’s A100.