AI BENCHMARK PROFILE

CANN Bench

General AICoding & Software Engineering

Proposes a benchmark for AI-generated operator code on Huawei's Ascend NPU, covering 53 operators and 1060 test cases. Evaluation uses a three-dimensional weighted composite score for compilation, functional correctness, and performance.

Released: 2026-07-08
Readiness: Paper only
Primary field: General AI

Why it matters

Could fill a gap in evaluating kernel-generation agents beyond CUDA/Triton, but clear artifact availability and a usable scoring contract are not confirmed from provided evidence.

Motivation

AI agents are now capable of writing, compiling, and iteratively optimizing low-level operator kernels on different hardware platforms.

Primary resources

Paper ↗Hugging Face ↗

Benchmark Radar records only publicly supported details and links back to primary sources for verification.