Huawei claims its AI chip is faster than the Nvidia A100 – but the Chinese superpower is struggling to match rivals’ pace of transformation as international sanctions force a design change
Nvidia cannot ship its high-end GPUs to China due to strict US export restrictions. Huawei has therefore done its best to fill the gap.
At the recent Nanjing World Semiconductor Conference, Wang Tao, Chief Operating Officer of Huawei’s Ascend and Kunpeng ecosystem, claimed that Huawei’s AI chip is comparable to, if not better than, Nvidia’s A100.
According to the South China Morning Post (SCMP)Wang stated that Huawei’s Ascend 910B AI chip delivers 80 percent of the efficiency of an Nvidia A100 when training LLMs, but “in some other tests, Ascend chips can beat the A100 by 20 percent.” Huawei first introduced the Ascend chip series in 2019, four months after the company was added to the US trading blacklist.
Change of design
Although Nvidia sells lower quality chips in China, it still claimed 90% of the AI processor market in 2023. By comparison, Huawei only had 6% of the market, but its share could and likely will improve.
While Wang maintained that there is “not much difference” between the Huawei 910B and the Nvidia A100 when training large AI models, a data snapshot from CSET (Center for Security and Emerging Technology) shows how US export controls are hampering Huawei production.
The report states: “This analysis of Huawei’s first generation Ascend 910 series (2019) and second generation Ascend 910B series (2022) suggests that despite export restrictions, Huawei was able to produce a better performing chip that could be produced domestically at SMIC (Semiconductor Manufacturing International Corporation).”
Despite that promising start, CSET adds: “Our analysis shows that the performance improvement is smaller than advertised; only 75 percent of the theoretical maximum performance improvement can be attributed to an actual increase in hardware performance. Additionally, Huawei has reduced the number of active AI cores between the 910 and 910B series – likely due to poor yields or limited capacity on SMIC’s 7nm manufacturing process.”
To compensate for the less active AI cores, Huawei added an additional vector unit in each core for a 25% performance boost and increased the clock speed, resulting in a 50% boost. As for the remaining 25% improvement, that appears to be the result of “a change in the way Huawei calculates peak performance.”
You can read CSETThe full findings here.