CANN Bench for Evaluating AI-Generated Kernels on Huawei Ascend NPUs
July 24, 2026
CANN Bench evaluates agent-written operator kernels on Huawei Ascend NPUs across 53 operators and 1,060 test cases. The benchmark measures compilation, functional correctness, and performance against hardware limits, covering precisions from FP32 to INT8.
HOW THIS AFFECTS YOU
●
builderYou can now benchmark agentic code generation specifically for the Ascend NPU ecosystem beyond CUDA.
●
researcherThis provides a principled reward signal for training agents to optimize low-level hardware kernels.