The chip contains 24 billion transistors and as well as performing graphics has been optimized for artificial intelligence training in the cloud and the company claims it achieves twice the performance of products in the market, making it a rival to Nvidia.
Big Island supports a variety of data types including: FP32, FP/BF16, INT32/16/8 and other multi-precision data and a single chip can perform 147 teraflops per second at FP16 resolution. The application programming interface supports mainstream deep learning development frameworks such as TensorFlow and PyTorch.
Design work on Big Island was started in 2018. Zheng Jinshan, co-founder and chief scientist of Tianshu Zhixin, said Big Island had taped out in May 2020 and come back from the foundry in November and was shown to operate in December 2020. Since then Big Island has been subjected to nearly 100 performance and validation tests.
The company said it would soon start mass production and commercial delivery of the product, one to two years ahead of its domestic competitors.
The company did not state which foundry was making the chip but the use of 7nm manufacturing process and 2.5D multi-die and interposer packaging labeled CoWoS (chip on wafer substrate) clearly point to Taiwan Semiconductor Manufacturing Co. Ltd. CoWoS is a TSMC name for 2.5D interposer technology.
Next: Chairman speaks