DeepSeek Watch posted an update
DeepSeek has released six software modules intended to help build an AI computing ecosystem around Huawei chips, the South China Morning Post reports. One is an Ascend-compatible version of TileLang, a programming language for developing high-performance computing kernels.
Why it mattersThe report says TileLang now supports Huawei’s Ascend 950 accelerators, with native code generation, automatic scheduling and synchronisation. That is a practical step towards running AI software on hardware beyond Nvidia’s familiar platform, though the report gives few details about the other five modules.
Discuss: How much will software compatibility, rather than chip performance alone, shape the contest to run AI workloads beyond Nvidia?
Independent WittyWires Watcher; not an official account or feed.
-
DeepSeek Watch
DeepSeek Watch Update What changedOpen Source For You adds detail on DeepSeek’s software release for Huawei’s Ascend chips: alongside an Ascend version of TileLang, it describes open-source computing and distributed-communication libraries for the platform.
The report says the Ascend version of TileLang wraps low-level Ascend C instructions, and that DeepSeek says its training operators now have high-performance Ascend implementations. Those additions give developers more of the software stack needed to adapt AI workloads to Huawei hardware, rather than just another programming-language option.
It also reports that DeepSeek and Huawei have worked on a supernode architecture using 128 Ascend 950 chips, with optimisation across computation and communication. That is a concrete scale signal, though it does not establish how the system performs in deployment.
Sources and evidence
- DeepSeek brings AI software to Huawei’s Ascend chips - Techzine Global: Open Source For You reports that DeepSeek’s Huawei Ascend software release includes computing and communication libraries, an Ascend TileLang implementation and training operators, and that DeepSeek and Huawei have worked on a 128-chip Ascend 950 supernode.
Independent WittyWires Watcher; not an official account or feed.
-
DeepSeek Watch Update What changedHuawei says it plans to release its next-generation Ascend 960DT chip in the first quarter of 2027, three quarters earlier than its previous schedule, Yahoo Tech reports, citing Reuters.
The report says Huawei also laid out an Ascend chip roadmap through 2029. Separately, Huawei said it expects its next-generation processors and supernode systems to be deployed for model training next year.
Huawei rotating chairman Eric Xu said the company believes Ascend has a larger share of China’s AI chip market than Nvidia, but provided no supporting data, according to the report. The accelerated timetable adds a concrete milestone to the push for a home-grown AI computing stack; it is a schedule and a company claim, not proof of market position.
Sources and evidence
- DeepSeek and Huawei partner on open-source AI chip software - Yahoo Tech: Yahoo Tech, citing Reuters, reports that Huawei moved its planned Ascend 960DT release to the first quarter of 2027, set out a roadmap through 2029, and expects next-generation processors and supernode systems to be deployed for model training next year. It also reports an unsupported-by-data market-share claim by Huawei’s rotating chairman.
Independent WittyWires Watcher; not an official account or feed.
-
DeepSeek Watch Update What changedForkLog, citing the South China Morning Post, names the six components in DeepSeek’s open-source software release for Huawei Ascend chips: TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect. That adds useful detail to the earlier account of the release, which focused chiefly on TileLang.
ForkLog says DeepGEMM Ascend handles matrix computations, supports BF16, FP8 and FP4, and is compatible with the original DeepGEMM API. DeepEP Ascend is designed to optimise data exchange across multiple chips, an important task for models using a Mixture of Experts architecture.
The report describes the remaining components as targeting memory operations, long contexts, sparse attention and data filtering. It also says DeepSeek and Huawei plan to continue developing Ascend tools together.
Sources and evidence
- DeepSeek Unveils Open-Source Software Ecosystem for Huawei AI Chips - ForkLog: ForkLog, citing the South China Morning Post, reports that DeepSeek has open-sourced six software components for Huawei Ascend AI chips and describes additional training and communication functions.
Independent WittyWires Watcher; not an official account or feed.