Huawei will reportedly start receiving samples of the Ascend 910D next month. The latter still has a ways to go before it is ready to be mass-produced for customers. Even at this early stage of development, Huawei is hoping that the chip can outperform Nvidia’s H100 which was first released in 2022 to train AI systems. Huawei has created some of the most popular substitutes for Nvidia’s AI semiconductors.
Not only is Huawei unable to be shipped cutting-edge chips but Chinese foundries are not allowed to purchase the Extreme Ultraviolet (EUV) lithography machines needed to build chips using a process node lower than 7nm. Nvidia’s H100 is built by TSMC using its 4nm node; in theory, this means that Nvidia’s AI silicon contains smaller transistors than Huawei’s allowing the American chipmaker to pack more inside a single chip. This makes Nvidia’s AI processors more powerful and energy-efficient.
Huawei’s new chip uses new packaging technologies that can integrate more silicon dies allowing Huawei to improve performance. SMIC, China’s largest foundry and the third largest in the world after TSMC and Samsung Foundry, is building the Ascend 910D. Because the foundry is banned from obtaining an EUV lithography machine, the chip must be made on SMIC’s 7nm N+2 process.
Not everyone believes that Huawei’s AI processors can rival the performance of such chips made by Nvidia. While the Ascend 910C was supposed to be a challenger to Nvidia’s H100, engineers who used both said that Huawei’s processors did not match the performance delivered by Nvidia’s AI chips.
