Neutrino - Fine-grained GPU Kernel Profiling via Programmable Probing

Learn about NEUTRINO, a programmable interface for GPU kernel profiling that achieves instruction-level granularity through assembly-layer probing in this 15-minute conference presentation from OSDI '25. Discover how this innovative tool overcomes limitations of existing GPU kernel profilers by providing hardware-independent, fine-grained measurements across both time and value domains. Explore the Densified Memory Access Timeline (DMAT), a novel visualization technique that reveals new insights into GPU runtime behavior. Understand the implementation details for both NVIDIA and AMD GPUs on Linux systems, and examine extensive evaluation results demonstrating NEUTRINO's superior profiling capabilities with minimal overhead. Gain insights into how this open-source tool can advance GPU performance analysis and optimization research in the era of scaling laws where understanding detailed GPU behavior is increasingly critical.