Nvidia is unable to ship its advanced GPUs to China due to strict US export restrictions, so Huawei has done its best to step in and fill the gap.
At the recent Nanjing World Semiconductor Conference, Wang Tao, COO of Huawei's Ascend and Kunpeng ecosystem, stated that Huawei's AI chip is on par, if not better, than Nvidia's A100.
According to the The South China Morning Post (SCMP), Wang claimed that Huawei's Ascend 910B AI chip offers 80 percent of the efficiency of an Nvidia A100 when training LLMs, but “in some other tests, Ascend chips can outperform the A100 by 20 percent.” . Huawei first introduced the Ascend series of chips in 2019, four months after the company was added to the US trade blacklist.
Change of design
Even though Nvidia sells lower-quality chips in China, it still holds 90% of the AI processor market in 2023. In comparison, Huawei had just 6% of the market, but the share could and likely will improve.
Although Wang insisted that there is “not much difference” between the Huawei 910B and the Nvidia A100 in training large AI models, a snapshot of data from CSET (Center for Security and Emerging Technology) shows how US export controls are hampering Huawei's production.
The report notes: “This analysis of Huawei's first-generation Ascend 910 series (2019) and second-generation Ascend 910B series (2022) suggests that despite facing export controls, Huawei was able to design a higher-quality chip. performance that could be manufactured domestically at SMIC (Semiconductor Manufacturing International Corporation).”
Despite that promising start, CSET adds: “Our analysis reveals that the performance increase is smaller than advertised; only 75 percent of the theoretical peak performance increase can be attributed to an actual increase in hardware performance. In addition, Huawei reduced the number of active AI cores between the 910 and 910B series, likely due to low yields or limited capacity in SMIC’s 7nm manufacturing process.”
To compensate for the lower number of active AI cores, Huawei added an additional vector unit in each core to increase performance by 25% and increased the clock speed, resulting in a 50% increase. As for the other 25% improvement, it appears to be the result of “a change in the way Huawei calculates maximum performance.”
You can read CSETThe full findings are here.