AscendGYM: A Cost-Aware Kernel Optimization Benchmark for Real-World Workloads on Ascend NPUs
AscendGYM measures kernel speedups against the token cost of achieving them.
The initial suite contains 48 kernels across three difficulty levels, with data types and tensor shapes drawn from real training and inference workloads. It tests optimization of existing Ascend NPU implementations in an environment designed to reproduce practical pipeline calls. Researchers can evaluate models with a built-in harness or test custom model-and-harness combinations. The proposed Speedup-Token Score weighs performance gains against token use to assess optimization within a budget.
HF Daily Papers' note
The initial suite contains 48 kernels across three difficulty levels, with data types and tensor shapes drawn from real training and inference workloads. It tests optimization of existing Ascend NPU implementations in an environment designed to reproduce practical pipeline calls. Researchers can evaluate models with a built-in harness or test custom model-and-harness combinations. The proposed Speedup-Token Score weighs performance gains against token use to assess optimization within a budget.
HF Daily Papers' note
score 5