Tianshu Zhixin debuts high-performance GPGPU

Tianshu Zhixin debuts China’s first domestically developed HPC GPU,
Image: Tianshu Zhixin

Shanghai Tianshu Zhixin Semiconductor Co., Ltd. (Tianshu Zhixin) recently announced it powered up China’s first domestically developed high-performance general-purpose graphics processor unit (GPGPU). The firm’s debut product has technical specifications that rival products made by Nvidia and Advanced Micro Devices (AMD).

The chipmaker’s GPGPU, dubbed Big Island, has incredible potential, but international tensions could affect its success.

Big Island’s Technical Specifications

Tianshu Zhixin revealed that Big Island features 24 billion transistors and offers 147 TFLOPS of FP16 computational power. The company’s chipset also supports several other floating-point formats, including BF16, FP32, INT8, INT16, and INT32.

The startup noted its component was fabricated using a 7nm node and boasts 2.5D chip-on-wafer-on-substrate (CoWoS) packaging. That means it has a relatively low power consumption rate, making it a cost-effective solution for web services providers.

Based on those specs, Big Island has a comparable density to Nvidia’s RTX 3080 and 3090 GPUs, which have 28 billion transistors. The processor’s computational capability edges out Nvidia’s A100, which can perform 77.97 TFLPOS at FP16. However, its latest data center product features an embedded tensor core that pushes its peak performance to 312TFLOPS.

In addition, Tianshu Zhixin’s GPGPU comes close to matching AMD’s Instinct M100’s capability of 184.6 TFlOPS at FP16.

The startup deserves praise for developing a product with world-class performance so quickly. The firm only began development on Big Island in 2018 and reached the tape out phase in May 2020. Its triumph reflects well on the entire Chinese electronic components industry.

Next Steps and Potential Headwinds

Tianshu Zhixin noted it is preparing to take its GPGPU into volume production to capitalize on its commercial applications. The firm said it completed almost 100 tests validating its chipset and would optimize its design with its partner companies. Its website lists Oracle, Dell, Inspur, and East Asian telecommunicates operator China Unicom as corporate associates.

Tianshu Zhixin explained Big Island would have applications in the education, autonomous vehicle, medical, and security fields, among others. The processor’s ability to execute various HPC functions and artificial intelligence (AI) training tasks would appeal to those sectors. The startup has not released a timeline for its first data center product’s official launch.

But it should see strong demand for its GPGPU from Chinese businesses and foreign companies working in the region.

Since 2018, the U.S. and China have been engaged in a trade war that has made exporting or importing electronic components from either country increasingly complicated. However, Tianshu Zhixin might avoid tariffs and trade restrictions emanating from the conflict as a Shanghai-based enterprise.

That said, the chipmaker may encounter headwinds in the future because of how it makes its GPGPU. Though the firm did not list a partner foundry, Big Island’s tech specs indicate Taiwanese Semiconductor Manufacturing Company (TSMC) fabricated it. The GPGPU is packaged using the CoWoS technique, which is a proprietary TSMC technology. Tianshu Zhixin also lists the contract fabricator as a business associate.

If Washington or Beijing issues new export controls that affect TSMC, Tianshu Zhixin could have trouble mass producing its processor. But right now, that issue is only theoretical, and its technological breakthrough is impressive. The global semiconductor industry should pay close attention to its next moves.


Please enter your comment!
Please enter your name here