Chinese-Company-To-Send-7nm-GPU-To-Production-Next-Quarter

The Company Is On Track To Send Its First 7nm GPU Design To Manufacturing In The Third Quarter Of The Year.

By Anton Shilov

Nvidia And AMD Continue To Dominate The GPU Market, But China-Based GPU Maker Biren Technology Hopes To Change That. The company is on track to send its first 7nm GPU design to manufacturing in the third quarter of the year; the company said in an interview (via CnTechPost). Biren Technology says that its compute GPU will compete with Nvidia’s next-gen 5nm compute GPUs (thought to be ‘Hopper’ and ‘Lovelace’).

Biren Technology’s initial product will be a compute GPU tailored specifically for AI workloads, so it’s designed for both training and inference. Biren will tape out the GPU in the third quarter this year (meaning it will send the design to manufacturing), and the chip will make its commercial debut sometime next year. Biren’s chip will come fabbed on the 7nm node, but it’s unclear whether the developer will use Taiwan-based TSMC or its rival, Korea-based Samsung Foundry.

Given the typical amount of time companies spend on bringing up their new silicon (assuming that it works as planned), it’s reasonable to expect Biren Technology to release its compute GPU commercially sometime in Q3 or Q4 2022.

Biren Technology received $170 million in Series A funding in 2020, so it does have at funding. Meanwhile, the development and implementation of a moderately complex FinFET-based 7 nm processor costs around $297.8 million, according to estimates by International Business Strategies (cited by SemiEngineering). A highly complex design can cost as much as $1.5 billion, IBS says. To that end, it remains to be seen how capable Biren’s AI chip will be and how complex it will be.

Interestingly, Biren CTO and Chief Architect Mike Hong said in an interview with Chinese media recently that Biren’s very first chip’s performance will rival Nvidia’s upcoming 5nm compute GPUs. Of course, being competitive doesn’t mean beating the Nvidia GPU, but rather co-existing on the market. Furthermore, there’s no way (short of industrial espionage) that Hong could know what Nvidia’s codenamed Hopper or Lovelace architectures are capable of, so we should take his comments with a grain of salt. 

Biren Technology was established in 2019 to develop GPUs suitable for AI, high-performance computing (HPC), and graphics rendering, a description on the company’s website reads. Biren does not plan to build a universal GPU immediately, so the first product will be designed just for cloud AI workloads (i.e., with a focus on INT8 and FP16 computing). Still, it will take advantage of ‘high-end packaging technology’ to address demanding applications. Eventually, Biren will move on to HPC applications and then will follow with a GPU that be used for AI, HPC, and graphics.

Biren will be the second Chinese company to use a 7nm node to build a GPU for AI and HPC applications, following Tianshu Zhixin with its Big Island GPU.

This news was originally published at Tomshard Ware.