China has taken another step on the way to becoming an independent semiconductor industry: According to its own statements, Shanghai Tianshu Zhixin Semiconductor has designed a GPU accelerator developed entirely at home, of which the first test chips are running in the laboratories. It is intended to be similar to Nvidia’s Ampere-GPU A 100 for data centers with a focus on machine learning, specifically for training AI algorithms. The actual GPU accelerator from Shanghai Tianshu Zhixin Semiconductor belongs to the “Big Island” family and includes 24 Billions of transistors, paired with four memory stacks of high-bandwidth memory (HBM, probably HBM2 or HBM2e). The company relies on a manufacturing process with structure widths of 7 nanometers – the packaging process “2.5D-CoWoS” speaks for contract work from TSMC.

China can apparently own Designing GPUs with an AI focus, but not yet producing them with modern exposure technology. The Chinese chip order manufacturer SMIC currently produces small quantities with 14 – nm structures.

High computing power On paper, the GPU accelerator presented can be displayed with 147 FP 16 – TeraFlops for a debut work. “Big Island” can also be used with the integer formats INT8, INT 16 and INT 32 and Bfloat 16 and FP 32 bypass. For comparison: Nvidia’s A 100 – card creates 78 FP 16 – TFlops over the shader cores and 312 FP 16 – TFlops using the integrated tensor cores (624 with sparsity, which removes zeros from the matrices). AMD’s Instinct MI 100 reaches around 185 FP 16 – TFlops.

The hardware is only the half the battle – the drivers, especially with GPUs, take up a lot of work. In the case of machine learning, the software stack is also important, where Nvidia and Co. have years of experience.

Huawei has with the Ascend 910 has already presented its own AI accelerator, but in the form of a processor without a graphics unit, which simplifies development. Before the trade war between China and the USA, Huawei had the chip manufactured by TSMC with 7 nm technology; meanwhile the manufacturer had to stop production. Alibaba, in turn, has its own 12 – Nanometer AI chip Huanguang 800 presented for the Alibaba cloud; But this is an AI inferencing processor.

(mma)