Shanghai Tianshu Zhixin Semiconductor Co., Ltd., generally referred to as Tianshu Zhixin, declared that its Big Island (BI) GPGPU has come to life. The BI is touted as the first domestic GPGPU in China that is targeted to AI and HPC applications and other sectors, such as education, medicine and defence.
24 billion transistors are bundled by the BI, and it is built on a home-made GPU architecture. The chip is designed with a cutting-edge 7nm process node and 2.5D CoWoS (chip-on-wafer-on-substrate) packaging. The foundry responsible for making the BI was not specifically revealed by Tianshu Zhixin. However, the description of the node coincides with one of TSMC’s manufacturing processes.
In 2018, the company committed to development on its BI chip. Back in May 2020, the firm finalised the tape out for BI and should already have undergone mass production if Tianshu Zhixin wants to achieve its goal of selling the chip this year.
Reportedly, the BI solution provides double with the efficiency of current mainstream goods on the market, while still offering a very appealing performance-to-cost ratio. BI supports a plethora of floating-point formats, including FP32, FP16, BF16, INT32, INT16, and INT8.
Tianshu Zhixin holds a tight lip on the performance of the BI, but the firm has teased up to 147 TFLOPS of FP16 performance. For comparison, the Nvidia A100 and AMD Instinct MI100 have FP16 performance estimates of up to 77.97 TFLOPS and 184.6 TFLOPS, respectively.