- DeepSeek released Ascend-optimized versions of its TileLang compiler, DeepGEMM, FlashMLA, TileKernel, DeepSelect operator libraries and DeepEP communication library.
- Benchmarks on Ascend 950 SuperPoD with EP32 deployment achieved 2,469 tokens/s per card at TPOT=5ms and 5,102 tokens/s at TPOT=10ms for DeepSeek-V4.1-Flash in offline mode.
- Huawei provided jointly defined SuperPoD Flex and UBL128 networking supporting 128-card 3.2Tbps Scale-up and 256K-card Scale-out configurations.
- All reference implementations and technical reports were published on Huawei's CANN open-source community.
- The release pairs a leading Chinese model vendor with domestic chip hardware, advancing China's independent AI software stack.
Open-Source Release Covers Full Software Stack
DeepSeek on September 30 formally open-sourced infrastructure components built for Huawei's Ascend computing platform, mirroring the GPU-oriented libraries it had previously released. The package includes TileLang, a high-level language compilation tool, alongside high-performance operator libraries DeepGEMM, FlashMLA, TileKernel and DeepSelect, plus the DeepEP distributed communication library.
Ascend provides open Ascend C API interfaces that enable developers to deeply optimize memory access paths and compute pipelines for high-performance operator development. The platform supports both fine-grained manual tuning by experienced developers and higher-level compilation through tools like TileLang.
Benchmarked Performance on Ascend 950
For large-scale inference scenarios using an EP32 deployment strategy, DeepSeek-V4.1-Flash running in offline inference mode without framework overhead achieved per-card output throughput of 2,469 tokens per second at a time-per-output-token (TPOT) of 5ms, and 5,102 tokens/s at TPOT=10ms on Ascend 950 SuperPoD clusters.
The benchmark conditions used a 128K context length with a Dspark speculative decoding acceptance rate of 0.85. The reported figures exclude serving-layer scheduling and framework load-balancing overhead, representing pure model performance.
The DeepEP communication library recorded measured interconnect bandwidth of 375 GB/s for Dispatch operations and 347 GB/s for Combine operations, which the teams described as approaching the hardware ceiling.
Joint Hardware-Software Collaboration
Huawei supplied jointly defined Ascend SuperPoD Flex and UBL128 networking architectures to the DeepSeek team, enabling 128-card 3.2Tbps single-layer Scale-up networks and 256K-card two-layer Scale-out networks designed to support ultra-low-latency inference and large-scale base model training.
Huawei published the resulting reference implementations and technical reports on its CANN open-source community, covering large-EP low-latency inference deployment, single-card and single-machine deployment, large-scale pre-training, long-context KV cache pooling, Agentic RL workflows and low-precision quantization training.
The collaboration represents a deepening integration between Chinese model developers and domestic chip vendors, providing the software tooling needed to make Ascend a viable alternative platform for frontier model training and deployment.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01
