Exjobbspresentation: Machine-Learning-Based Correction of Analytical NPU Performance Models

Modern Neural Processing Units (NPUs) must be evaluated long before silicon exists, yet early tools force a difficult compromise: fast analytical performance models enable exploration, but may diverge from slower RTL simulation, which provides a more accurate timing reference.This thesis investigates whether that gap can be learned and corrected. Using early-available block-command descriptors, sequence summaries, and analytical cycle estimates from Arm’s NPU modelling flow, super-vised XGBoost regressors are trained as a correction layer rather than a replacement for the existing model. The method is evaluated on anonymised single-block
and short-sequence workloads across three NPU units, comparing corrected predictions against RTL cycle measurements with a 5% acceptance criterion. Results show large improvements in difficult cases, raising several low baseline pass rates above 90% while substantially reducing median and tail errors. SHAP-based analysis further identifies workload regimes linked to systematic mismatch, making the correction model not only more accurate but also a practical diagnostic tool for model improvement.

Handledare: Jonas Skeppstedt

Examinator: Pietro Andreani