Ascend 950
An AI chip product in Huawei's Ascend series, designed for large-scale AI training and inference cluster scenarios.
Core Facts
Timeline (last 90 days)
The open-source version of DeepInfer covers only one model (DeepSeek V4.1 Flash) and two types of hardware (NVIDIA SM100 and Ascend 950), and supports only offline inference; KV Cache storage, PD disaggregation, load balancing, and online serving are all not included.
Currently the hardware supported by DeepSeek-Infer is NVIDIA SM100 GPUs and Huawei Ascend 950.
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.
DeepInfer supports two types of hardware: NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs.
The performance of the Ascend 950's 128-card super-node solution has already approached the hardware ceiling.
The Ascend 950 is Huawei's latest-generation AI training chip, and the 128-card super-node solution refers to connecting 128 Ascend 950 chips via high-speed interconnects to form a single logical training unit.
DeepEP Ascend achieves a sustained Dispatch bandwidth of 90%~95% of the physical payload limit on the Ascend 950, on the condition that the EP scale does not exceed 32; larger EP scales and Combine are still being optimized.
The Huawei team provided strong support, and both parties jointly advanced the 128-card super-node solution based on the Ascend 950.
DeepGEMM Ascend was first released on 2026.09.30, supports Ascend 950 devices, is fully compatible with DeepGEMM's API, supports BF16, FP8, and FP4 matrix multiplication, and uses the same package name as the original version.
FlashMLA's prefill throughput for the sparse attention operator on the Ascend 950 reaches 410 TFLOPS (95% of the hardware's theoretical peak), and decode reaches 360 TFLOPS (83%), both figures as published by DeepSeek itself with no third-party measurements.
4 more timeline events
All Facts (14)
DeepInfer supports two types of hardware: NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs.
70%UnverifiedThe open-source version of DeepInfer covers only one model (DeepSeek V4.1 Flash) and two types of hardware (NVIDIA SM100 and Ascend 950), and supports only offline inference; KV Cache storage, PD disaggregation, load balancing, and online serving are all not included.
50%UnverifiedCurrently the hardware supported by DeepSeek-Infer is NVIDIA SM100 GPUs and Huawei Ascend 950.
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%UnverifiedThe performance of the Ascend 950's 128-card super-node solution has already approached the hardware ceiling.
50%UnverifiedThe Ascend 950 is Huawei's latest-generation AI training chip, and the 128-card super-node solution refers to connecting 128 Ascend 950 chips via high-speed interconnects to form a single logical training unit.
50%UnverifiedDeepEP Ascend achieves a sustained Dispatch bandwidth of 90%~95% of the physical payload limit on the Ascend 950, on the condition that the EP scale does not exceed 32; larger EP scales and Combine are still being optimized.
50%UnverifiedFlashMLA's prefill throughput for the sparse attention operator on the Ascend 950 reaches 410 TFLOPS (95% of the hardware's theoretical peak), and decode reaches 360 TFLOPS (83%), both figures as published by DeepSeek itself with no third-party measurements.
50%UnverifiedDeepSeek and Huawei jointly optimized the Ascend 950 large-scale cluster solution, performing in-depth tuning for multi-card training and large model inference.
50%UnverifiedThe Ascend 950 single cabinet is equipped with 64 cards and 3,192 chips, the majority of which are switching chips responsible for managing inter-card data communication.
50%UnverifiedDeepGEMM Ascend was first released on 2026.09.30, supports Ascend 950 devices, is fully compatible with DeepGEMM's API, supports BF16, FP8, and FP4 matrix multiplication, and uses the same package name as the original version.
50%UnverifiedThe Huawei team provided strong support, and both parties jointly advanced the 128-card super-node solution based on the Ascend 950.
50%UnverifiedHuawei has submitted a bid for an AI data center project to the Egyptian government, proposing to export at least 1,408 Ascend 950 series chips for AI training, along with an additional 600 chips for building an inference cluster.
50%UnverifiedHuawei's Ascend 950 super-node physical machine has officially made its debut, achieving 1,024-card interconnection with computing power reaching the exaflops scale.
50%