DeepSeek V4 is compatible with Huawei Ascend, igniting the computing power sector

1、 Market explosion: Liyang chip leads gains, computing power sector collectively strengthens

On the morning of April 13th, the concept of computing power chips quickly rose, and Liyang chips were the first to hit the 20% daily limit up, becoming the core leading indicator of the sector's growth. Driven by it, Wantong Development simultaneously hit the daily limit up, with domestic computing related stocks such as Haiguang Information and Loongson Zhongke leading the way in gains, and funds pouring into the domestic chip market. This market trend is not simply about hype, but a positive response from the market to a major breakthrough in the domestic AI ecosystem - the deep adaptation between DeepSeek V4 and Huawei Ascend has broken the long-term dependence of high-end models on overseas chip and software ecosystems, allowing the market to see the full chain autonomous and controllable potential of domestic computing power from hardware to software, and thus igniting the enthusiasm of funds for the layout of related industry chain enterprises.

2、 Technological breakthrough: full stack deep adaptation, building a solid foundation for domestic AI computing power

The adaptation of DeepSeek V4 to Huawei Ascend this time is not a simple port compatibility, but a full chain deep optimization from the underlying architecture, model design to software stack. Both teams spent several months completing the comprehensive migration from overseas CUDA framework to Huawei CANN framework, rewriting over a million lines of underlying code, and conducting specialized optimizations for the operators, quantization, and cluster scheduling of Ascend chips. After adaptation, the model's inference speed is significantly improved, the utilization rate of computing power is significantly increased, and the entire process from training to inference is autonomously controllable, completely eliminating dependence on overseas high-end chips.

As the world's first top-level large model to run entirely on Huawei Ascend chips, the successful adaptation of DeepSeek V4 has verified the ability of domestic chips to carry trillion parameter level large models. The Huawei Ascend 950PR chip provides hardware support for efficient operation of large models with its powerful computing power and storage performance, while the DeepSeek V4's technological strength fully unleashes the computing power value of domestic chips. The two work together to build a complete independent closed loop of "domestic chips+top large models", filling the core gap of China's AI industry's "chip shortage and soul shortage".

3、 Industrial significance: Breaking the ecological monopoly and ushering in a new era of domestic AI independence

For a long time, the global AI chip market has been monopolized by overseas manufacturers, and the barrier of high-end training chips and supporting software ecology has become a key bottleneck restricting the development of China's AI industry. The deep adaptation between DeepSeek V4 and Huawei Ascend not only tears open the gap of overseas ecological monopoly, but also opens up a new era of independent development of domestic AI industry.

From an industrial perspective, this breakthrough will accelerate the maturity of the domestic computing power industry chain: upstream chip design and manufacturing enterprises will usher in a broader market space, midstream testing and packaging demand will continue to be released, and downstream AI application enterprises will also receive more stable and secure computing power support. At the same time, this successful case will drive more large model manufacturers and hardware companies to join the domestic ecosystem construction, promoting the transformation of domestic AI technology from "single point breakthrough" to "ecological rise".