The growing symbiosis indicates that DeepSeek may primarily rely on Huawei’s already mature technology stack by 2027 or 2028, SemiAnalysis said in its report.
Deepening “symbiosis” between Chinese artificial intelligence model startup DeepSeek and Chinese chipmaking giant Huawei is closing the technological gap with US top technology firm Nvidia. By harnessing their respective strengths, both the companies are fuelling domestic innovation, according to a report by the research firm SemiAnalysis.
DeepSeek has unveiled software tools where Huawei’s powerful Ascend 950PR chips are increasingly experimented with. One of the primary ones is the Ascend-compatible version of TileLang, a custom programming language designed to streamline development of high-performance kernels (essential code for operating chip hardware).
This growing symbiosis indicates that DeepSeek may primarily rely on Huawei’s already mature technology stack by 2027 or 2028, SemiAnalysis said in its report. This aligns with Beijing’s strategy of technological self-reliance since the US imposed export restrictions of high-performance Nvidia chips to China.
“From the software results so far and testing on 950PR, the ecosystem gap is definitely closer with DeepSeek’s help on top of Huawei’s work. Jensen’s (Huang) nightmare is starting to come true,” SemiAnalysis said in its report.
DeepSeek has also provided Huawei with early access to its models like V4 and V4.1 for “Day 0 inference support”. Through this support Huawei deploys its chips to run the AI models and not just train them.
The Ascend 950PR chip, unveiled by Huawei in March, lags Nvidia’s H100 – which the company has been shipping since 2022- in absolute performance according to the report. Moreover Ascend’s hardware configuration outside DeepSeek’s supernode set-up still remains a challenge. Developers have faced issues trying to run Ascend hardware on standard x86 host processors instead of native Huawei architectures.
However, Ascend kernels are achieving high levels of efficiency at 99.7 per cent of their theoretical ceiling. Moreover other operators serving DeepSeek can reuse the kernels, while other labs can build new models on shared compiler capabilities. If the issues can be fixed, over the course of time, the gains of the continued collaboration between the two companies can compound according to the report
