Huawei Ascend
Huawei's AI chip and computing platform brand, targeting large-model training and inference scenarios. It provides a full-stack AI solution spanning chip hardware and software toolchains.
Core Facts
Cube Studio explicitly supports Huawei Ascend NPUs and domestic CPU/GPU adaptation.
DeepSeek has open-sourced infrastructure components for the Ascend platform, covering the TileLang compiler, compute library, and communication library.
On September 30, 2025, DeepSeek open-sourced a batch of underlying infrastructure components, including operators, a programming language, compute libraries, and communication libraries
Timeline (last 90 days)
vLLM-Ascend is vLLM's backend adaptation for Huawei Ascend NPUs, enabling models to run on non-GPU domestic accelerator cards.
In April 2026, the DeepSeek V4 preview was released, including Flash with 284 billion parameters and Pro with 1.6 trillion parameters, supporting million-level context, all under the MIT license, with Huawei Ascend integrating support at the first opportunity.
DeepInfer is released under the Apache 2.0 license, supports NVIDIA Blackwell and Huawei Ascend 950, and includes components such as expert load balancing, speculative decoding, and asynchronous scheduling.
According to the video information, DeepInfer is compatible with both the NVIDIA and Huawei Ascend platforms, a single cluster can reach a daily throughput of 155 trillion tokens, and each GPU can earn $38 per hour.
主流NPU被广泛应用于手机、自动驾驶和边缘服务器中，如华为昇腾、苹果神经引擎和高通的Hexagon处理器。
An NPU (neural processing unit) is a processor designed specifically for neural network inference and training tasks, and its architecture typically includes large multiply-accumulate arrays (MAC arrays), on-chip caches, and dedicated data scheduling units.
Huawei Ascend introduced the AscendC programming framework in its early years.
Where Huawei Ascend hardware truly lags behind NVIDIA is in software ecosystem; Huawei has built its own CANN ecosystem but still has gaps when it comes to supporting mainstream large models.
DeepSeek open-sourced DeepGEMM Ascend edition, a BLAS-level GEMM operator library specifically designed to run on Huawei Ascend hardware.
DeepGEMM-Ascend's target is explicitly and exclusively Huawei Ascend, and it does not establish a universal software layer applicable to all chip vendors.
21 more timeline events
All Facts (20)
Cube Studio explicitly supports Huawei Ascend NPUs and domestic CPU/GPU adaptation.
80%VerifiedDeepSeek has open-sourced infrastructure components for the Ascend platform, covering the TileLang compiler, compute library, and communication library.
70%VerifiedOn September 30, 2025, DeepSeek open-sourced a batch of underlying infrastructure components, including operators, a programming language, compute libraries, and communication libraries
65%Unverified美国对华芯片管制客观上刺激了中国自主芯片研发的加速推进，华为昇腾系列芯片、寒武纪等国产AI芯片研发明显提速
80%Partially VerifiedAn NPU (neural processing unit) is a processor designed specifically for neural network inference and training tasks, and its architecture typically includes large multiply-accumulate arrays (MAC arrays), on-chip caches, and dedicated data scheduling units.
65%UnverifiedAccording to the video information, DeepInfer is compatible with both the NVIDIA and Huawei Ascend platforms, a single cluster can reach a daily throughput of 155 trillion tokens, and each GPU can earn $38 per hour.
60%UnverifiedvLLM-Ascend is vLLM's backend adaptation for Huawei Ascend NPUs, enabling models to run on non-GPU domestic accelerator cards.
50%UnverifiedDeepInfer is released under the Apache 2.0 license, supports NVIDIA Blackwell and Huawei Ascend 950, and includes components such as expert load balancing, speculative decoding, and asynchronous scheduling.
50%Unverified主流NPU被广泛应用于手机、自动驾驶和边缘服务器中，如华为昇腾、苹果神经引擎和高通的Hexagon处理器。
50%UnverifiedHuawei Ascend introduced the AscendC programming framework in its early years.
50%UnverifiedWhere Huawei Ascend hardware truly lags behind NVIDIA is in software ecosystem; Huawei has built its own CANN ecosystem but still has gaps when it comes to supporting mainstream large models.
50%UnverifiedThe original DeepGEMM is a CUDA kernel library, and the Ascend version provides a relatively familiar interface for Ascend while handling layout and data movement.
50%UnverifiedDeepGEMM-Ascend's target is explicitly and exclusively Huawei Ascend, and it does not establish a universal software layer applicable to all chip vendors.
50%UnverifiedThe interconnect bandwidth of DeepSeek operators on Huawei Ascend supernodes has approached the hardware limit.
50%UnverifiedTileLang was open-sourced and adapted for Huawei Ascend chips at the first opportunity.
50%UnverifiedHuawei Ascend has its own software stack, CANN, but its ecosystem richness is far inferior to that of CUDA.
50%UnverifiedTileLang allows developers to write a single codebase that can be compiled and run on both NVIDIA servers and Ascend servers, without complex Ascend adaptation.
50%UnverifiedMoE places extremely stringent demands on multi-card communication, memory bandwidth, and operator scheduling, and general-purpose CUDA operator libraries often cannot directly bring out the best performance of Ascend hardware.
50%UnverifiedWhat has been open-sourced this time are operators and libraries, not models, which does not mean Ascend can directly run DeepSeek services.
50%UnverifiedThe open-sourced Ascend components cover operator programming languages, matrix computation, multi-card communication, and long-context processing.
50%