Ascend 950PR
A Huawei Ascend-series AI inference accelerator card equipped with 128 GB of memory. It supports low-precision inference formats including MXFP4 and MXFP8, and is designed for large-model inference deployment scenarios requiring high throughput and memory capacity.
Timeline (last 90 days)
The Huawei Ascend 950PR successfully ran a pruned version of the DeepSeek V4-Flash model in a single-card configuration, achieving a peak decode speed of 208 TPS.
The Ascend 950PR is equipped with 128 GB of VRAM.
The code related to this benchmark test has been open-sourced.
With an approximately 1.6M ultra-long context and 128 concurrent requests, throughput can reach approximately 3,700 tokens/s (aggregated).
The content creator used the npu-top tool together with Pin Agent to have the local model autonomously complete adaptation for the 950PR; after enabling acceleration, a speed of approximately 6× was achieved, and the model completed all adaptation tasks within approximately 6 minutes.
Running higher-specification models on the 950PR is currently slower than on the Pro7000 platform, and deep optimization has not yet been performed.
These test data come from preliminary tests by a single source in a personal short-term usage environment (several hundred hours of 950PR usage time) and do not constitute a standardized benchmark.
All Facts (5)
The Huawei Ascend 950PR successfully ran a pruned version of the DeepSeek V4-Flash model in a single-card configuration, achieving a peak decode speed of 208 TPS.
50%UnverifiedThe Ascend 950PR is equipped with 128 GB of VRAM.
50%UnverifiedWith an approximately 1.6M ultra-long context and 128 concurrent requests, throughput can reach approximately 3,700 tokens/s (aggregated).
50%UnverifiedThe code related to this benchmark test has been open-sourced.
50%UnverifiedThe content creator used the npu-top tool together with Pin Agent to have the local model autonomously complete adaptation for the 950PR; after enabling acceleration, a speed of approximately 6× was achieved, and the model completed all adaptation tasks within approximately 6 minutes.
50%