A linear layer at batch 1 runs at a third of a percent of peak and at batch 4,096 near 70%. The curve between has two regions and a sawtooth: the memory-bound slope where intensity equals M, the plateau where the tensor cores are the limit, and the dips where the tile count does not divide the SM count.
Updated Sep 2026 · Grounded in real AI infrastructure interview loops and written to a senior-engineer editorial bar, with every number worked and every diagram hand-built.
A linear layer at batch 1 runs at a third of a percent of peak and at batch 4,096 near 70%. The curve between has two regions and a sawtooth: the memory-bound slope where intensity equals M, the plateau where the tensor cores are the limit, and the dips where the tile count does not divide the SM count.
Ranked by how closely each one overlaps this question's topic, so the first card is the thing to read if the answer above moved too fast.
Scored on deriving intensity as a function of M, placing the ridge on the batch axis, and explaining wave quantization with the tile and SM counts rather than as a vague 'under-utilization'.
No comments yet — be the first to share your approach.
