Tuesday, November 17, 2026
AI 인프라 · 뉴스 & 분석
홈 › 반도체·하드웨어 › 리포트
반도체·하드웨어 · 리포트

DeepSeek와 Huawei는 Huawei의 Ascend 950 AI 가속기를 위한 오픈소스 프로그래밍 도구 및 라이브러리를 공개했으며, Nvidia 에코시스템 락인을 줄이기 위한 컴퓨트 및 통신 라이브러리를 포함합니다.

Ascend 오픈소스 소프트웨어 생태계는 개발자의 전환 비용을 낮추고 훈련·추론용 Ascend 950 가속기의 광범위한 도입을 가능하게 하며 Nvidia 스택 외 AI 칩 독립성을 진전시킵니다.
업계 전문지Slicast · 2026년 10월 1일 14:00 UTC · 글로벌 · 출처: Tom's Hardware
중요도 78

DeepSeek has released open-source programming tools developed with Huawei for the company's Ascend AI chips, aiming to reduce reliance on Nvidia's software and hardware ecosystem. According to a September 30 Reuters report, the release includes open-source libraries for AI computation and chip-to-chip communication, along with native Ascend support for the high-level programming language TileLang. DeepSeek says these tools—developed with full support from Huawei—are intended to simplify programming and enable developers to fully leverage the hardware's performance.

The companies optimized computation and communication across a supernode system built around 128 Ascend 950 chips. Their work addresses two core requirements for running large AI workloads across multiple accelerators: performing calculations efficiently on each chip and moving data quickly between them to keep the processors occupied.

The released libraries include DeepGEMM-Ascend, which handles matrix multiplication and other calculations used in DeepSeek's models. It supports BF16, FP8, and FP4 operations and uses the same programming interfaces as DeepSeek's existing DeepGEMM library, enabling developers to retain familiar APIs when moving to Ascend. DeepEP-Ascend handles communication for training and inference, including routing data to different experts in mixture-of-experts models and combining their outputs. Both libraries were developed and tested on Ascend 950 hardware.

TileLang provides the higher-level programming layer for writing optimized kernels. DeepSeek described it as offering "a simpler programming model" than Nvidia's CUDA, aimed at improving development efficiency and simplifying code. The language already supported Nvidia and other hardware, with earlier adapters available for Huawei's Ascend processors. The September 30 update adds native support for Ascend 950, including code generation, automatic scheduling, and synchronization.

The tools build on Huawei's existing CANN software platform, which provides the underlying infrastructure for running AI workloads on Ascend. Nvidia's CUDA platform has long supplied developers with a mature programming environment and libraries optimized for its GPUs, making the software ecosystem a major part of the company's advantage in AI computing. While DeepSeek's release provides developers with additional tools to optimize workloads on Huawei hardware, TileLang's support for multiple platforms ensures the language remains useful for Nvidia GPUs.

The announcement comes two weeks after Huawei unveiled its next generation of AI processors and supernode systems. Huawei expects its AI systems to be widely used for model training in 2027. DeepSeek and Huawei had already collaborated on DeepSeek's V4 model, released in preview form in April, with support for Huawei's Ascend chips. Huawei confirmed that its Ascend 950 supernodes fully support the V4 models and that its chips were used for part of the lighter V4-Flash model's training.

원문 보기
DeepSeek와 Huawei는 Huawei의 Ascend 950 AI 가속기를… · Slicast