DrivingSDK

Analyze Ascend NPU profiling data to identify training performance bottlenecks. Breaks down step-level time into compute, unoverlapped communication, and freetime; within compute, analyzes compute vs memory-bound ratios and cube vs vector utilization to summarize the model's performance bottleneck.

Predict cross-chip performance from Ascend NPU profiling data. Invoke when user provides a profiling directory path and asks to predict/estimate performance on other chips.