AI Architecture & Systems

Roofline and Performance Upper Bounds

1.13.1Arithmetic Intensity and Data Traffic at Each Memory Boundary#

1.13.2Compute Roof, Memory Roof, and Ridge Point#

1.13.3Peak Roofline and Empirical Roofline#

1.13.4Hierarchical Register/SRAM/DRAM Roofline#

1.13.5Batch Size, Matrix Shape, and Reuse in GEMM/GEMV#

1.13.6Beyond the Bandwidth Bound: Launch Latency, Synchronization, Dependencies, and Utilization Limits#

1.13.7Data-Movement Energy and the Energy Roofline#