Chinese artificial intelligence company DeepSeek joined forces with Huawei on September 30 to develop programming tools tailored for Huawei’s Ascend AI chips. This alliance addresses China’s mounting demand for alternatives to Nvidia through the use of open-source software and finely tuned libraries built for Ascend processors.
Making the announcement via WeChat, DeepSeek noted that Huawei backed the creation of the Ascend programming infrastructure. Additionally, the partners progressed on a 128-chip Ascend 950 supernode, refining both computing capabilities and communication processes for massive AI workloads.
As part of this open-source rollout, the companies introduced TileLang, a high-level programming language engineered specifically for AI accelerator creation. According to DeepSeek, TileLang streamlines code logic and boosts development productivity across various chip platforms.
DeepSeek additionally deployed Ascend-compatible versions of several essential software components relied upon in AI processing. This lineup features DeepGEMM for matrix computations, DeepEP for device-to-device communication, TileKernels for vector processing, FlashMLA for sparse attention, and DeepSelect for data filtering.
Rather than depending entirely on low-level chip tools, this software initiative establishes a comprehensive development layer surrounding Huawei’s Ascend hardware. DeepSeek stated that every TileLang operator utilized in the training of its V4-series now features a high-performance Ascend version.
Developers can utilize identical Python interfaces for both Nvidia GPUs and Huawei NPUs, with the system automatically determining the appropriate backend. This level of compatibility aims to reduce friction for engineering teams transitioning workloads between distinct AI accelerator environments.
Furthermore, DeepSeek framed TileLang as an alternative to Nvidia’s CUDA software framework. “TileLang was created precisely to meet this need,” DeepSeek explained, pointing to the demand for a less complicated programming model for AI chips.
This collaboration emerged shortly after Huawei unveiled its newest generation of AI processors alongside supernode architectures. Huawei anticipates that these systems will see more widespread adoption for model training in the coming year.
This initiative bolsters the software foundation supporting China’s larger drive toward domestic AI computing infrastructure. By leveraging open-source tools, developers can construct applications around Ascend hardware while cutting down on their reliance on Nvidia’s well-established software ecosystem.
Ultimately, the partnership equips Huawei’s AI processors with a broader developer framework geared toward large-scale model training. It also reinforces the software ties linking domestic hardware accelerators with DeepSeek’s expanding ecosystem of AI models.
Also Read: NVIDIA Introduces OpenShell, Sentry to Strengthen AI Agent Safety
