DeepSeek and Huawei release open‑source tools to program Ascend 950 and cut reliance on CUDA

DeepSeek announced on September 30, 2026 the public availability of several open‑source programming tools, developed in collaboration with Huawei, for the Ascend 950 accelerators. The initiative is presented as a strategic step to reduce the friction developers experience when they rely exclusively on Nvidia’s CUDA ecosystem, and to broaden the range of hardware options available to the artificial‑intelligence community. The announcement also highlights the two companies’ commitment to open innovation and to encouraging the adoption of diversified hardware solutions.

Context of the AI ecosystem

For many years, CUDA has become the de‑facto platform for training and inference of large‑scale models, thanks to its maturity, the abundance of optimized libraries, and the widespread availability of Nvidia GPUs in data centers. This dominance has created a structural dependency that makes it difficult to adopt alternative architectures, even when those alternatives offer cost or energy‑efficiency advantages. The entrenched position of CUDA in common workflows means that teams face a substantial effort to rewrite and validate code for other platforms. Consequently, the community is actively seeking alternatives that can match performance while offering greater flexibility.

The new tools from DeepSeek and Huawei

DeepGEMM‑Ascend is the first of the released libraries. It is a matrix‑multiplication layer that supports BF16, FP8 and FP4 precisions while maintaining familiar interfaces for users already working with the original DeepGEMM version. The intention is to enable developers to migrate their code with minimal changes, leveraging the Ascend 950 chip architecture for both training and inference workloads. This interface compatibility aims to cut adaptation time and simplify integration into existing pipelines.

• BF16
• FP8
• FP4

DeepEP‑Ascend complements DeepGEMM by handling device‑to‑device communication during distributed training and large‑scale inference. The library includes routines for routing mixture‑of‑experts models, a technique that distributes parts of a model across different chips to improve efficiency. According to statements from both companies, the solution is designed to integrate with the usual PyTorch and TensorFlow workflows, providing a smooth transition for teams already invested in those frameworks.

TileLang, Huawei’s high‑level programming language, received native support for Ascend 950 in this release. The extension enables automatic code generation, automatic task scheduling, and operation synchronization, thereby reducing the manual effort normally required to optimize for specific hardware. TileLang’s cross‑platform support indicates that users can continue working with other architectures without abandoning Nvidia, although the release does not claim performance parity.

Architecture and testing on a 128‑chip supernode

To validate the performance of the new libraries, DeepSeek and Huawei ran tests on a supernode composed of 128 Ascend 950 chips. The results showed improvements in compute efficiency and communication when DeepGEMM‑Ascend and DeepEP‑Ascend were used together, thanks to integration with Huawei’s CANN platform, which manages resource allocation and hardware‑level task orchestration. These tests demonstrated the system’s ability to fully exploit the parallelism offered by a large number of chips.

Despite the advances, the statements make clear that the release does not equate to abandoning Nvidia nor achieving performance parity with CUDA. The maturity of the CUDA ecosystem remains a competitive advantage, and the published benchmarks do not cover all workload types or network configurations. Consequently, adopting Ascend 950 will require organization‑specific evaluations to determine whether the solution meets their particular needs.

Implications for developers and data centers in Latin America

For AI development teams in the region, the availability of DeepGEMM‑Ascend, DeepEP‑Ascend and TileLang support opens the possibility to experiment with alternative hardware without abandoning their PyTorch or TensorFlow‑based workflows. Data centers that already have Huawei infrastructure can now leverage a more complete software stack, which could translate into lower licensing costs and greater supplier diversification.

In practice, organizations that decide to try Ascend 950 accelerators will need to review their training pipelines to integrate the new libraries, adapt their distributed‑communication scripts, and validate performance on their own data sets. Huawei’s expectation of broad system use for training throughout 2027 suggests that support and documentation will continue to evolve, easing the transition for those who adopt the solution in the coming years.