DeepSeek Opens Huawei Ascend Support for Core AI Libraries

Overview

In a significant move toward diversifying its hardware dependencies, DeepSeek has open-sourced versions of its core AI training and inference libraries compatible with Huawei’s Ascend architecture. This development effectively reduces the company’s reliance on NVIDIA hardware, marking a strategic pivot in the rapidly evolving landscape of open-source artificial intelligence.

The release includes four critical components: DeepGEMM, DeepEP, TileLang, and FlashMLA. These libraries form the foundational software stack that powers DeepSeek’s high-performance models. By making these tools available for Ascend chips, the company is providing the broader developer community with the means to train and serve models on non-NVIDIA hardware without sacrificing performance.

Strategic Shift in Hardware Ecosystem

The announcement signals a deliberate effort by DeepSeek to decouple its technological progress from the NVIDIA supply chain. For years, the AI industry has been heavily reliant on NVIDIA GPUs, particularly for high-performance computing tasks. However, supply constraints, geopolitical tensions, and escalating costs have prompted many leading AI labs to explore alternative hardware solutions.

DeepSeek’s decision to prioritize Huawei Ascend support addresses these challenges directly. The Ascend 950 series, Huawei’s flagship AI processor, offers a viable alternative to NVIDIA’s latest offerings. By ensuring compatibility at the library level, DeepSeek is lowering the barrier to entry for other researchers and organizations looking to adopt Huawei hardware.

This move also aligns with a growing trend in the open-source AI community. As models become more complex and resource-intensive, the ability to optimize software for diverse hardware architectures is becoming a key competitive advantage. DeepSeek’s contribution helps democratize access to high-performance AI infrastructure, reducing the monopoly of any single hardware provider.

Technical Breakthroughs and Performance Metrics

Among the released libraries, DeepGEMM stands out as the most critical component. It serves as the matrix-math engine underlying every DeepSeek model, handling the fundamental linear algebra operations required for both training and inference. Optimizing this engine for a new architecture is a significant engineering achievement.

The Ascend port of DeepGEMM is designed to maintain API compatibility with the NVIDIA version. This design choice ensures that developers can switch hardware backends with minimal code modifications, preserving workflow continuity. Furthermore, the library supports multiple precision formats, including BF16, FP8, and FP4, which are essential for maximizing throughput and minimizing memory usage on modern AI chips.

Performance benchmarks provided by Huawei indicate strong results. In offline tests using the V4.1-Flash configuration, the system achieved a throughput of 2,469 output tokens per second per card. This metric highlights the efficiency of the optimized kernels, demonstrating that Huawei hardware can deliver competitive performance for large-scale language model serving.

Implications for Model Serving Costs

Faster computational kernels directly translate to lower operational costs. When each chip can process more tokens per second, the overall infrastructure required to serve a model decreases. This efficiency is crucial for the sustainability of open-source AI projects, which often operate on tight budgets.

By enabling high-performance serving on Ascend hardware, DeepSeek is helping to drive down the cost of running sophisticated AI models. This trend toward cheaper model serving benefits everyone in the ecosystem, from independent developers to large enterprises. It also encourages further innovation in hardware-software co-design, as vendors compete to provide the best performance per watt and per dollar.

Looking Ahead

The open-sourcing of these libraries is expected to accelerate adoption of Huawei Ascend hardware within the global AI research community. As more developers test and refine these tools, feedback loops will likely lead to further optimizations and stability improvements.

This development does not necessarily signal the end of NVIDIA’s dominance in the AI hardware market. However, it firmly establishes Huawei as a credible alternative for serious AI workloads. For DeepSeek, this strategy ensures resilience against supply chain disruptions and positions the company as a leader in hardware-agnostic AI engineering.

As the field of artificial intelligence continues to mature, the ability to build efficient, scalable systems on diverse hardware platforms will be increasingly important. DeepSeek’s latest contribution sets a precedent for how open-source models can thrive in a multi-vendor hardware environment.