What DeepSeek’s Ascend AI Stack Does—and What Its Benchmarks Show

DeepSeek’s September 30, 2026 release added Ascend versions of six components spanning kernel programming, compute, attention, data selection and distributed communication. The software provides building blocks for running AI workloads on Huawei hardware, but the available evidence does not show a like-for-like performance match with NVIDIA—or confirm that DeepSeek completed full-scale V4 training on Ascend. 25

What DeepSeek released

The stack pairs components with roles in DeepSeek’s NVIDIA-oriented software, while using Ascend-specific implementations where needed. 25

• TileLang is the high-level kernel programming and compilation tool. Its Ascend support provides a way to develop kernels for Huawei hardware without implying that the kernels are identical to NVIDIA versions. 25
• DeepGEMM is a matrix-multiplication library. Its Ascend port keeps the DeepGEMM API and supports BF16, FP8 and FP4 operations, as well as MQA logits and MegaMoE. The project describes Ascend-specific optimizations, including sparse data loading and coroutine-based pipelining. 17
• TileKernels supplies general-purpose kernels; FlashMLA provides sparse-attention prefill and decoding kernels for Ascend 950; and DeepSelect handles data selection or filtering. 91444
• DeepEP handles distributed communication, including expert-parallel dispatch and combine operations. Its public buffer APIs are aligned with the NVIDIA version, while its Ascend implementation uses Ascend-specific software and hardware interfaces. 19

That alignment is useful for developers, but API or component correspondence is not proof that the two platforms have identical internals, full feature parity or comparable speed.

What the Ascend 950 numbers say

A reported DeepEP test on Ascend hardware measured 375 GB/s for dispatch and 347 GB/s for combine. Those figures describe Ascend communication performance; they are not a head-to-head comparison with NVIDIA. 1947

For inference, a reported DeepSeek-V4.1-Flash test used an Ascend 950 UBL128 system, an EP32 deployment strategy, a 128K context and offline inference. At a 5 ms time per output token (TPOT) setting, it reported 2,469 output tokens per second per card; at 10 ms TPOT, it reported 5,102 tokens per second per card. The reported figures excluded serving-scheduling and framework load-balancing overhead. 47

These are configuration-specific results, not a general throughput guarantee. The DeepEP repository describes its measurements as using a manually configured proof-of-concept hardware development package; another report says the figures used PoC firmware and pointed to a commercial release planned for around October 15. That means the firmware and setup matter when interpreting or trying to reproduce the numbers. 1219

The supplied reporting does not provide enough detail to independently reproduce every inference setting, and it describes no large-scale independent third-party validation. The results therefore should not be treated as a verified comparison against NVIDIA systems. 47

Huawei’s SuperPoD Flex and UBL128 disclosures

Huawei and DeepSeek described a jointly defined Ascend 950 SuperPoD Flex / UBL128 design. Reports give the scale-up configuration as 128 cards with a 3.2 Tbps single-stage network, and describe a two-stage scale-out design intended to extend to 256,000 cards. That is a stated design capability, not evidence that a system of that size ran a DeepSeek training job. 104749

Huawei also said joint work would be shared through the CANN community, with reported coverage including model deployment, large-scale training, long-context KV-cache pooling and agentic reinforcement learning. 14

What remains unproven

The release establishes a broader Ascend software stack, and the reported inference run shows a specific offline configuration. But neither component correspondence nor an advertised scale-out design confirms a full-scale V4 training run on Ascend. The cited benchmark figures also do not show NVIDIA-equivalent performance: they use different hardware, and the available reports do not provide a matched cross-platform test. 251947

The clearest reading is that DeepSeek has published meaningful Ascend software and Huawei has described a system architecture and selected test results. The extent to which those results translate to other workloads—or compare with NVIDIA in a controlled test—remains unresolved.