AI Architecture & Systems

Profiling, Benchmarking, and Correctness

SourcesProfiling tools in the corpus layer mappings; benchmark fundamentals retained from the original outline.

2.19.1Microbenchmarks, Warm-up, Asynchronous Timing, and Timing Boundaries#

2.19.2Nsight Compute/Systems and ROCm Profiling#

2.19.3Device-Specific Analysis Tools: Neuron Explorer, PopVision, and Compiler Reports#

2.19.4Roofline, Stall Reasons, Memory Efficiency, and Communication Occupancy#

2.19.5Dense Peak, Sparse Peak, Achieved FLOPS, and Effective Model Work#

2.19.6Numerical Tolerance, Gradient Checking, Low-Precision Error, and Differential Testing#

2.19.7Race Detection, Regression Testing, and Cross-Device Performance Portability#