Authors: Mengyue Huang, Zhanhui Li, Tim Purcell, Mingqiao Li, Zhengru Wang, and Nirmal Saxena (NVIDIA Corporation)
Abstract: Accurate kernel-level GPU performance projection is essential for early-stage architecture design when target hardware is unavailable and only high-level specifications are known. Existing methods face a trade-off: simulators reveal bottlenecks but are slow and require unavailable hardware details, while analytical and end-to-end learning models are efficient yet rely on idealized assumptions and provide limited explainability.
We present DeepSim, an explainable GPU performance projection framework for single-GPU kernels that learns unit-level speed-of-light (SOL) ratios to calibrate bottleneck pressure using multi-task learning. Given target chip specifications and kernel traffic from existing GPUs, DeepSim constructs idealized limiters, predicts per-unit SOL ratios, and maps them to practical limiters reflecting target-chip behavior. Using this representation, DeepSim infers dominant bottlenecks, classifies workloads, and estimates execution time. Across 290K samples and 284 kernels spanning NVIDIA Ampere, Ada, Hopper, and Blackwell GPUs, DeepSim achieves 12.29% mean absolute percentage error on held-out B200 and 14.78% on unseen kernels.
Presentation: file