DeepSeek and Huawei release open-source Ascend AI tools to reduce Nvidia reliance
Building the Future: How DeepSeek and Huawei Are Rewriting the Rules of AI Programming
The race for cutting-edge artificial intelligence is increasingly defined not just by the power of the chips, but by the software that governs them. In a significant move to decentralize the AI landscape and reduce reliance on established ecosystems, Chinese AI powerhouse DeepSeek has joined forces with Huawei to develop revolutionary open-source programming tools for their Ascend AI chips.
This collaboration is more than just a partnership; it’s an ambitious effort to unlock the full potential of specialized hardware. By creating these tools, the two companies are aiming to simplify the complex task of programming AI workloads, allowing developers to maximize performance on Huawei’s hardware while operating outside of the dominant software frameworks.
At the core of this innovation is a focus on efficiency. The goal is to ensure that massive AI calculations are performed with maximum speed, and that data movement between processing units is virtually instantaneous. This addresses the fundamental challenge of running large AI models across multiple accelerators, demanding both efficient local computation and rapid communication.
The resulting suite of open-source libraries tackles these challenges head-on. DeepSeek released DeepGEMM-Ascend, a library designed to handle intensive matrix multiplications and calculations essential for AI models, supporting common data formats like BF16, FP8, and FP4. Crucially, these libraries maintain familiar programming interfaces, allowing developers to leverage existing knowledge while utilizing the unique capabilities of the Ascend hardware.
To handle the complex data flow required for training and inference, DeepEP-Ascend provides robust communication management. This tool expertly routes data to different components within mixture-of-experts models and efficiently combines their results, ensuring that even the most intricate AI architectures run smoothly across the supernode system built around 128 Ascend 950 chips.
Perhaps the most transformative component is TileLang, the high-level programming layer. DeepSeek positioned TileLang as a simpler alternative to established frameworks like Nvidia’s CUDA. By introducing native support for the Ascend 950, TileLang brings powerful features like automatic code generation, scheduling, and synchronization directly to the hardware, dramatically simplifying the development process.
This new layer builds directly upon Huawei’s existing CANN software platform, providing the necessary infrastructure for running advanced AI workloads. While the broader AI industry has long been anchored by Nvidia’s mature software ecosystem, this development offers developers a compelling alternative that prioritizes hardware optimization and platform independence.
This synergy positions the partnership at the forefront of the evolving AI hardware narrative. As Huawei continues to push the boundaries of its AI processors and supernode systems, these open-source tools ensure that the hardware is matched by equally powerful, accessible software, fostering an environment where innovation can thrive beyond existing boundaries.
Buy on Amazon
Credit
