FlashMLA Hits 95% Theoretical Speed on Ascend 950

Newly open-sourced DeepSeek libraries report FlashMLA reaching 95% of theoretical performance on Huawei's Ascend 950, with DeepSelect up to 20x faster.

@sharedot

FlashMLA Hits 95% Theoretical Speed on Ascend 950

By @sharedot · · 8 pages

  • Programming
  • Deepseek
  • Huawei Ascend
  • Tilelang
  • Open Source

Newly open-sourced DeepSeek libraries report FlashMLA reaching 95% of theoretical performance on Huawei's Ascend 950, with DeepSelect up to 20x faster.

The benchmarks buried in the release

The headline news — six open-source tools for Huawei Ascend chips — was already widely reported, but the performance claims inside the release are the story working programmers will chase.

According to GIGAZINE, FlashMLA, the library that accelerates attention processing in DeepSeek-V4.1 inference, achieved 95% of its theoretical performance on the Ascend 950 for some processes, per GitHub reporting.

GIGAZINE also reports that DeepSelect, which speeds the TopK selection used in DeepSeek Sparse Attention, can accelerate the standard torch.topk by 2 to 20 times depending on conditions.

What the six modules actually do

The release mirrors DeepSeek's existing Nvidia tooling. TileLang now officially supports Ascend 950 with native code generation, automatic scheduling, and synchronization, and Bloomberg called it "China's answer to Nvidia CUDA," per ForkLog.

DeepGEMM Ascend handles matrix computations with the same API as the original and supports BF16, FP8, and FP4, with kernels designed to operate near the hardware's performance limits, ForkLog reports.

DeepEP Ascend optimizes chip-to-chip data exchange for Mixture of Experts architectures, while the remaining components cover memory operations, long contexts, sparse attention, and data filtering.

The 128-chip supernode angle

DeepSeek and Huawei jointly optimized computation and communication on a "supernode" built from 128 Ascend 950 chips, according to Tom's Hardware.

The work targets the two core requirements of large AI clusters: calculating efficiently on each chip and moving data fast enough between accelerators to keep processors occupied.

Tom's Hardware notes both DeepGEMM-Ascend and DeepEP-Ascend were developed and tested on Ascend 950 hardware, with the libraries built on Huawei's existing CANN software platform.

Why the stakes go beyond Huawei

Nvidia's CUDA ecosystem has an estimated four million developers, according to The New York Times reporting cited by The Next Web, and that software moat is widely seen as the hardest part of Nvidia's advantage to replace.

Tom's Hardware points out that TileLang retains multi-platform support, including Nvidia and AMD hardware, so the language remains useful even for developers who never touch Ascend.

DeepSeek said the Ascend version of TileLang is designed as an example of building a highly accessible software ecosystem for a wider range of AI chips, per ForkLog.

Scale: 160,000 Ascend chips planned

According to Bloomberg, as reported by The Next Web, DeepSeek plans to deploy at least 160,000 Huawei accelerators in a data center in Inner Mongolia — a scale that explains why the software stack matters as much as the silicon.

The Next Web also reports DeepSeek is raising funds at a valuation close to 500bn yuan ($74bn), and that DeepSeek's annualized revenue run rate reached $1bn last week.

Seoul Economic Daily notes DeepSeek is nevertheless said to use Huawei chips for inference while still relying on Nvidia chips to train its models.

What comes next

The companies plan to continue joint development of tools for Ascend, ForkLog reports, and Huawei's next-generation Ascend 960DT chips are slated for the first quarter of 2027 — an earlier date than initially scheduled, according to ForkLog.

Tom's Hardware notes the collaboration builds on prior work: Huawei said its Ascend 950 supernodes fully supported DeepSeek's V4 models and its chips were used for part of the lighter V4-Flash model's training.

Sources

  1. gigazine.net › DeepSeek has partnered with Huawei to open-source six AI development tools
  2. forklog.com › DeepSeek Unveils Open-Source Software Ecosystem for Huawei AI Chips
  3. tomshardware.com › DeepSeek and Huawei release open-source Ascend AI programming tools
  4. thenextweb.com › DeepSeek open-sources Huawei Ascend tools as a CUDA alternative
  5. en.sedaily.com › DeepSeek Teams With Huawei to Challenge Nvidia's CUDA

More on Programming