AsianFin -- China’s state media in a report earlier this week suggested Alibaba Group has developed a new artificial intelligence (AI) processor that is set to challenge Nvidia Corporation’s H20, the most powerful AI chip Nvidia is created to comply with the Biden-era export controls and allowed to sell in China.
Credit：Alibaba
The report, aired Tuesday on China Central Television (CCTV), involves Chinese Premier Li Qiang’s visit to China Unicom’s Sanjiangyuan Energy Intelligent Computing Centre in Qinghai province. Multiple Chinese companies including Alibaba’s semiconductor design unit T-Head were reported to have signed or plan to sign contracts for the computing power of the Unicom’s project.
A screenshot from the broadcasting showed a table comparing key parameters of the T-Head’s PPU, and application-specific integrated circuit, with Nvidia’s H20 and A800 GPU, or graphics processing unit, Huawei Ascend 910B processor and Shanghai-based Biren Technology’s 104P. The PPU leads in most of these metrics comparing with domestic competitors.
The T-Head’s PPU appears more powerful than Nvidia’s A800 and comparable with H20. The chip features 96 gigabytes (GB) of high-bandwith memory 2 extended (HBM2e) per unit, matching Nvidia H20's memory capacity and exceeding Nvidia A800. the PPU card boasts chip-to-chip interconnect bandwidth of 700 GB per second, surpassing Nvidia A880’s bandwidth of 400 GB/s and just shy of Nvidia’s H20. In terms of energy efficiency, the PPU card has a 400-watt power consumption, the same level with Nvidia’s A800, which is lower than the 550-watt that the H20 required.
News about Alibaba’s AI chips swirled these days, part of latest signs that Chinese tech giants are ramping up efforts to help ease their demands for Nvidia’s products amid trade tensions between China and U.S.
Alibaba, a long-time big customer of U.S. AI chip leader Nvidia, has developed a new AI chip "that is more versatile than its older chips", and the company and other chip designers could begin "filling the void left after Nvidia ran into regulatory barriers to selling its products in China," the Wall Street Journal reported on August 29.
Unlike previous cloud-computing chips mainly designed for specific applications, Alibaba’s new chip, now in testing, appeared as a general-purpose one since it was reported to serve a broader range of AI interference task. The chip is designed for inference, instead of training, as China’s biggest weakness is training AI models, the report noted.
Alibaba’s new chip was also said to be designed to work with the Nvidia software ecosystem so it will be compatible with the Nvidia platform to ease engineers and developers’ adoption. The report suggested the chip is domestically built. That was an essential step for China toward more localized production, though local players relying on Chinese chip factories like Alibaba are now facing the challenge to get enough supply due to manufacturers’ difficulties in increasing capacity.
Alibaba and Baidu Corporation have been using their self-designed chips to train their AI models, partly replacing Nvidia chips, The Information reported last Thursday, noting that neither of their chips are comparable with Nvidia’s most advanced products.
Alibaba has reportedly started using its own chips, the Zhenwu perse processing unit, for training its smaller models since early this year, and the chip is good enough as a rival of Nvidia’s H20. Alibaba is also working on a more advanced version of Zhenwu, which is set to slightly outperform Nvidia’s A100, the data center GPU for AI rolled out five year ago, Per the report. That was seen as a sign of how far behind the company’s chips are.
