Ascend NPU
A series of AI-dedicated processors (NPUs) developed by Huawei for deep learning training and inference workloads. The series is paired with the CANN software stack and the HCCL communication library, and constitutes a major domestic AI computing platform in China.
Timeline (last 90 days)
Fused operators improve hardware utilization by running Cube and Vector compute units in parallel (CV parallelism) and reusing on-chip data to reduce memory read/write overhead.
The compute units of Ascend NPU are divided into Cube units (responsible for high-parallelism operations such as matrix multiplication) and Vector units (responsible for element-wise operations such as activation functions, normalization, and quantization), which are physically independent and can execute in parallel.
All Facts (2)
Fused operators improve hardware utilization by running Cube and Vector compute units in parallel (CV parallelism) and reusing on-chip data to reduce memory read/write overhead.
50%UnverifiedThe compute units of Ascend NPU are divided into Cube units (responsible for high-parallelism operations such as matrix multiplication) and Vector units (responsible for element-wise operations such as activation functions, normalization, and quantization), which are physically independent and can execute in parallel.
50%