Chinese chips: SMIC does not have enough 7-nanometer capacity for Huawei

Huawei has a new AI accelerator in its range, the Ascend 910B. The Chinese chip contract manufacturer SMIC apparently produces it using 7-nanometer technology. However, Huawei itself has not officially unveiled it, so there are no details on the specifications.

TSMC manufactured the previous Ascend 910 for Huawei, but the chip contract manufacturer is no longer allowed to produce for Huawei due to trade restrictions. The previous Ascend 910 is a chiplet design that splits the logic into two chiplets and also uses four HBM2E memory stacks. Two dummy chiplets stabilize the design.

It remains to be seen whether the successor will also consist of several silicon components or a monolithic chip. SMIC itself cannot assemble multi-chiplet processors, but its partner JCET can according to its website for so-called advanced packaging.

Huawei must prioritize

Meanwhile, the news agency Reuters is reporting production bottlenecks because SMIC is apparently unable to expose enough wafers. Until now, SMIC has produced the Kirin 9000S smartphone processor for Huawei's Mate 60 Pro using 7 nm technology. The partners have reportedly decided to reduce smartphone production in favor of AI accelerators. However, the quantities are not known.

The global market leader for AI accelerators is Nvidia with its H100. AMD is now trying to catch up with the MI300X and MI300A. In order to compete with these products from a speed perspective, Huawei must have significantly accelerated the Ascend 910B compared to the previous 910. The latter manages 256 trillion arithmetic operations with 16-bit floating point values per second (256 FP16 teraflops) and 512 INT8 teraOPS with eight-bit integers.

Nominally, Nvidia's H100 and AMD's MI300X accelerators are considerably faster. Nvidia, for example, claims just under one FP16 petaflops (1000 teraflops) when the Tensor cores are used. With a computing trick that omits the zeros in the matrices, the computing power theoretically doubles. AMD quotes 1.3 FP16 petaflops for its MI300X. However, the figures are only comparable to a limited extent, especially for AI accelerators. Huawei's Ascend 910B is said to be fast enough to have aroused Baidu's interest, among others.

AMD and Nvidia are not allowed to sell their top AI models to China due to export restrictions. Nvidia is therefore said to have launched another slower version, the H20, which is around half as fast as the H100. Previously, the A800 and H800 came with throttled interconnects in order to circumvent the US sanctions. The USA then imposed new rules.

(mma)