A Chinese tech company has just released a graphics processing unit (GPU) that it claims has world-beating computing abilities.
The state-of-the-art processor will be used for cloud-based AI training in data centres and has already attracted interest from several server providers.
A product of Shanghai-based Biren Technology, the BR100 seven-nanometre chipset was launched mid-August and reportedly performs three times better than its market peers.
The GPU was capable of performing a quadrillion floating-point operations per second (1 petaFLOP), said Biren chief executive, Zhang Wen – in comparison, a brand new Nvidia RTX 3080 GPU used in high-end gaming computers has around 30 trillion FLOPS (30 teraFLOPS).
Under normal circumstances, GPUs are equipped to perform more intensive tasks than central processing units (CPUs), whose function is to execute basic instructions. GPUs achieve their processing power by running multiple processors per chipset.
Biren Technology’s BR100 GPU is being advertised as a “general-purpose” chip, designed for use beyond creating graphical fidelity and high-frame rates in consumer electronics.
The BR100 chipset will see much of its use in large-scale cloud based-AI training, said Zhang.
Biren’s focus was on “accelerated computing scenarios” in data centres through partnerships with prominent Chinese tech firms like Inspur, one of the country’s largest server providers, according to a company statement.
Zhang added the BR100 showed promise when it came to processing a data format used in AI learning called TF32+-.
Chip manufacturers would all prosper from a quote “unified ecosystem” where users were free to create and develop applications, said Tsinghua University electrical engineering department head, Professor Wang Yu.
The BR100s impressive specs can be attributed to its innovative design and its seamless compatibility with other new technologies.
Biren hired former Nvidia Shanghai general manager, Yang Chaoyuan, earlier in the year. Yang, a 35-year industry veteran, is credited with improving the firm’s back-end supply chains.
Relying on performance figures alone may not be enough to convince industry players of the product’s maturity, said Bu Rixin, president of Chuangdao Consulting.
For example, Nvidia’s new four-nanometre GPU possesses 80 trillion transistors whereas Biren’s example houses 77 trillion. Biren’s chipset has a significantly larger surface area, measuring 1 000mm² compared to Nvidia’s 840mm².
Whether the BR100 GPU finds commercial success remains to be seen. The chip market is at the centre of a trade war between the US and China.
According to SCMP, Biren’s new product will be manufactured in Taiwanese chip foundries and is at risk of sanction.