is reinventing its AI chips from scratch
Cambricon Technologies is an essential company in China’s plans to challenge the US for its leadership in artificial intelligence (AI). Although it is not as well known as Huawei or Moore Threads, this is one of the companies specialized in the design of accelerators for AI with greater growth potential. Be that as it may, these three companies are China’s clearest alternatives to Nvidia because all three have already managed to place competitive solutions on the market. The priority strategy of the Government led by Xi Jinping seeks to build a self-sufficient ecosystem capable of breaking Nvidia’s dominance in the market. However, as stated SCMPat the center of this rivalry is a fundamental design debate: should China continue betting on GPUs or is it preferable for it to make the leap towards ASIC technology (Application-Specific Integrated Circuit or application-specific integrated circuit)? ASIC chips are designed to perform a single specific task, unlike GPUs and CPUs, which are general purpose. Its main advantage is efficiency. And since they are optimized for a specific function, they consume less energy and are faster in that task. Even so, they have a disadvantage: their rigidity. They cannot be reprogrammed to carry out another function, so the debate we raised a few lines above makes perfect sense. Convergence seems inevitable Large Chinese technology companies that choose ASIC chips for AI gain performance in their specific models, but are tied to an architecture that does not adapt well if the type of workload changes. This is the problem with this approach. a report prepared by Morgan Stanley and published on May 8, makes the market dynamics clear: it predicts that Huawei will capture 62% of the Chinese AI accelerator market in 2026, followed by Cambricon Technologies with 14%. ASIC chip heavyweights increasing relevance and volume in China Among the large technology companies with their own chips, Baidu and Alibaba are around 5% each. In any case, there is no doubt about one thing: the heavyweights of ASIC chips are increasing in relevance and its volume in China. And they are largely succeeding because the performance gap between Chinese chips and Nvidia GPUs allowed for export has narrowed noticeably. Morgan Stanley data reflect that Huawei’s Ascend 950 cards and Cambricon Technologies’ Siyuan 690 cards exceed the performance of Nvidia’s H20 GPU. Zhang Haijun, an expert semiconductor analyst, holds that as AI models become more complex the line between custom ASICs and flexible GPUs becomes increasingly blurred. This scenario suggests that the winning architecture could end up combining elements of both approaches. Su Lian Jye, the chief analyst at consulting firm Omdia, defend That companies with strong AI engineering capabilities and a clear roadmap benefit from ASICs, while those handling mixed workloads continue to lean toward general-purpose GPUs. For now, the market momentum in China clearly favors specialists. To companies that bet on ASIC technology. Partly by choice. Partly because the sanctions have left them no choice. Image | Enflame More information | SCMP In Xataka | The US remains committed to stopping China. Now it has targeted the second largest Chinese chip manufacturer
