Tianshu Zhixin Semiconductor, a fabless joint venture between Via Technologies and the Shanghai municipal government, has revealed what appears to be China’s first 7nm GPGPU, a massive chip designed to compete in the data centre with the likes of Nvidia A100 and AMD MI100.
China is pursuing “technological self-sufficiency,” and the trade war with the United States has intensified the process. China is capable of covering more than 20% of the chips used in the domestic market, but the government intends to raise that figure to 70% by 2025. One example is the use of ageing chips as the country’s semiconductor industry catches up, with the country turning to Japan for etching equipment that would otherwise be deemed outdated.
Last year, we saw the arrival of an amazing x86 CPU from the government-backed Zhaoxin, demonstrating just how far we’ve come in that regard. We’ve also learned that Huawei is hard at work developing server GPUs, but things have been pretty quiet after the tech behemoth lost access to main suppliers like TSMC.
Tianshu Zhixin was believed to be working on the first general-purpose GPU constructed in China, codenamed “Big Island.” The business was only able to do a paper launch in March due to the current processor scarcity, so what’s fascinating about these chips is that they’re supposedly capable of competing with Nvidia and AMD’s GPU at the top end.
Big Island GPUs aren’t technically intended to compete with RTX and Radeon gaming graphics cards, but rather for machine learning, high performance computing, medical science, and security. This means they’re supposed to compete with Nvidia’s A100 and AMD’s Instinct MI100, all of which are massive chips tailored for the data centre and give orders of magnitude more efficiency than previous generation architectures while taking up much less room and power to run.
Big Island was created between 2018 and 2020, and it was based on TSMC’s 7nm process node and 2.5D CoWoS packaging. They also make use of TSMC’s 65 nm silicon interposer and have 24 billion transistors. Tianshu Zhixin, like AMD and Nvidia, has fitted its GPGPU with 32 GB of HBM2 memory (1.2 TB per second bandwidth) and made it PCIe 4.0 compatible.
The Chinese GPGPU’s claimed success appears to be impressive. When it comes to FP16 (half precision math) efficiency, it can achieve 147 teraflops, which is comparable to the Nvidia A100’s 78 teraflops and the AMD MI100’s 184.6 teraflops – but it should be remembered that the A100’s Tensor cores can achieve 312 FP16 teraflops.
Big Island should be capable of up to 37 teraflops for FP32 workloads, which is more than the A100 and MI100, which give 19.5 teraflops and 23.1 teraflops, respectively. The new GPU’s power consumption is estimated at 300 watts, and Tianshu Zhixin claims it will have a better price-to-performance ratio than rival solutions.
Time can say if this is valid, and if China has reached a significant milestone in its efforts to reduce its dependence on international semiconductors.
110 Interactions, 8 today