AscendGYM Benchmark for Ascend NPU Kernel Optimization
November 16, 2026
AscendGYM evaluates how well models and agents optimize 48 different kernels using real-world tensor shapes and data types from actual workloads. It supports both model and model-harness evaluation modes for NPU training and inference efficiency.
HOW THIS AFFECTS YOU
●
builderYou can use this to test if an agent can actually improve kernel performance on Ascend hardware.