Ascend
A series of AI chips for training and inference introduced by Huawei. It is a principal representative of domestically produced AI computing in China and competes with NVIDIA GPUs in the domain of large-model training.
Core Facts
TileLang has been validated and proven mature on the NVIDIA platform, and carries the implementation of most operators in the training of the DeepSeek V4 series models.
DeepSeek has open-sourced infrastructure components for the Ascend platform, covering the TileLang compiler, compute library, and communication library.
Timeline (last 90 days)
On September 30, DeepSeek open-sourced a complete set of foundational Ascend components benchmarked against NVIDIA CUDA, comprising a programming language and five core compute components.
The Ascend ecosystem previously lacked tools similar to Triton, causing the barrier associated with hand-written TIK/AscendC to remain high.
The Ascend version of DeepEP uses HCCL, with UB within a node, HCCL/HCCN's URMA engine across nodes, and kernels written in Ascend C.
The Ascend version of DeepEP must use the expanded layout (Dispatch Expand) and currently operates as a single communication domain.
DeepSeek plans to deploy more than 160,000 Huawei Ascend AI accelerator chips at its Inner Mongolia data center.
Both parties have established a network topology for single-layer switching across 128 cards based on Ascend hardware, and claim it is scalable to larger clusters.
From the same TileLang code, the compiler can generate both NVIDIA's PTX instructions and Huawei Ascend's low-level instructions, demonstrating hardware neutrality.
TileLang has been validated and proven mature on the NVIDIA platform, and carries the implementation of most operators in the training of the DeepSeek V4 series models.
On September 30, DeepSeek announced the open-sourcing of a full set of underlying software for Huawei's Ascend platform, corresponding one-to-one with the suite previously built for NVIDIA chips
Ascend NPU compute cores are divided into three categories: AIC (matrix compute core / Cube core), AIV (vector compute core), and AICPU (scalar control logic)
7 more timeline events
All Facts (17)
TileLang has been validated and proven mature on the NVIDIA platform, and carries the implementation of most operators in the training of the DeepSeek V4 series models.
65%VerifiedDeepSeek has open-sourced infrastructure components for the Ascend platform, covering the TileLang compiler, compute library, and communication library.
70%UnverifiedHuawei's Ascend series AI chips still lag behind NVIDIA's top products in absolute performance, but have already achieved large-scale deployment in the domestic market.
60%UnverifiedThe Ascend ecosystem previously lacked tools similar to Triton, causing the barrier associated with hand-written TIK/AscendC to remain high.
50%UnverifiedThe Ascend version of DeepEP must use the expanded layout (Dispatch Expand) and currently operates as a single communication domain.
50%UnverifiedThe Ascend version of DeepEP uses HCCL, with UB within a node, HCCL/HCCN's URMA engine across nodes, and kernels written in Ascend C.
50%UnverifiedFrom the same TileLang code, the compiler can generate both NVIDIA's PTX instructions and Huawei Ascend's low-level instructions, demonstrating hardware neutrality.
50%UnverifiedDeepSeek plans to deploy more than 160,000 Huawei Ascend AI accelerator chips at its Inner Mongolia data center.
50%UnverifiedBoth parties have established a network topology for single-layer switching across 128 cards based on Ascend hardware, and claim it is scalable to larger clusters.
50%UnverifiedUB (Unified Bus) is a high-speed inter-chip interconnect technology unique to the Ascend platform, serving a similar role to NVIDIA's NVLink but implemented differently
50%UnverifiedAscend NPU compute cores are divided into three categories: AIC (matrix compute core / Cube core), AIV (vector compute core), and AICPU (scalar control logic)
50%UnverifiedAscend is Huawei's AI accelerator chip series, with the Ascend 950 being its latest high-end model, positioned to compete with NVIDIA's H100/H200.
50%UnverifiedXingchen 4.0 29B has completed end-to-end adaptation for multiple domestically produced chips including Ascend based on Zhiyuan FlagOS, enabling cross-chip architecture migration and one-click deployment.
50%UnverifiedSome Chinese enterprises adopt a dual-stack strategy: using NVIDIA GPUs during the research and development phase, then adapting trained models for Ascend inference deployment.
50%UnverifiedOn September 30, DeepSeek open-sourced a complete set of foundational Ascend components benchmarked against NVIDIA CUDA, comprising a programming language and five core compute components.
50%UnverifiedOn September 30, DeepSeek announced the open-sourcing of a full set of underlying software for Huawei's Ascend platform, corresponding one-to-one with the suite previously built for NVIDIA chips
50%UnverifiedThe DeepSeek community has open-sourced two underlying acceleration projects targeting the Ascend platform: DeepEP Ascend and DeepGEMM Ascend
50%