A look at Huawei and DeepSeek's mutual reliance, where Huawei provides compute hardware while DeepSeek contributes model workloads and kernel optimization

You can now use more powerful AI models on Chinese-developed hardware.

Huawei and DeepSeek have formed a symbiotic relationship where Huawei provides its Ascend AI hardware and software stack, while DeepSeek contributes its advanced AI models and kernel optimization expertise. This collaboration addresses their individual limitations: Huawei lacks a frontier lab to test its chips with real-world models, and DeepSeek needs access to the necessary compute power. The recent release of Ascend support for DeepSeek's tools like TileLang and DeepGEMM, as of September 30, exemplifies this partnership. Huawei's Ascend 950PR chip, designed for prefill and recommendation tasks, shows competitive memory bandwidth utilization, nearly matching Nvidia's H100 in some metrics, though it falls behind in absolute compute performance compared to high-end GPUs like A100 and MI300X. Huawei also employs a binning strategy for its chips, creating multiple SKUs by pairing compute and memory based on die quality. DeepSeek's software engineering is adapting to Huawei's hardware, for instance, by co-locating quantized values and scaling factors in memory to reduce data transfer overhead.

This partnership signifies a strategic move towards a more self-sufficient Chinese AI ecosystem, driven by necessity and mutual benefit. Huawei's Ascend hardware, while not leading in absolute performance, is becoming viable for advanced AI workloads due to DeepSeek's software optimizations. The development suggests that specialized hardware with optimized software can bridge performance gaps, making domestic solutions competitive for specific market needs, particularly in China where ecosystem building is a priority.

The integration of DeepSeek's kernel optimizations into Huawei's Ascend platform highlights the critical role of software in unlocking hardware potential. This collaboration will likely foster further innovation in AI hardware-software co-design within China, aiming for greater efficiency and performance for domestic AI development. Future efforts will focus on refining these integrations, potentially leading to more tailored solutions and a stronger position for Chinese AI technology globally.