The Ascend optimization benchmark

AI-powered kernels.
Measured performance.

Explore how AI agents optimize Ascend operators. Compare real speedups, inspect every method, and trace the code behind the results.

Code
baseline ÷ optimized = speedupAbove 1.0× is faster. Success threshold: > 1.05×.
ASCEND / COMPUTE ENGINEBENCHMARK
REPRODUCIBLECode → Run → Compare

01 / PERFORMANCETop operator speedups

The fastest operator–method pairs in your current selection.

Weighted speedup

02 / DISTRIBUTIONSpeedup distribution

Operator counts by measured speedup. Unavailable results are excluded.

03 / EFFICIENCYOptimization yield

Successful operators by final trial budget · speedup > 1.05.

Input + output tokens, including cache

04 / COMPARISONModel comparison

成功率按当前筛选的算子统计;平均 Token 用量与费用按有记录的试验计算,包含优化未成功的试验。DeepSeek 费用区间的下限为全部按空闲时段价格计算,上限为全部按高峰时段价格计算;高峰时段为北京时间周一至周五 09:00–12:00、14:00–18:00,其余时间为空闲时段。

05 / EXPLOREROperator explorer

Search operators and compare methods across repositories and categories.

loading

Select an operator name for case details and code diffs. Click a method heading to sort; scroll the table to compare all methods.

Latest updates

Product and benchmark snapshot history.

—
Benchmark snapshot loaded
Performance is calculated from recorded baseline and optimized runs. Token usage and cost estimates show their recorded coverage.