BLOG ILLUSTRATION / EXECUTION STORY

Accumulate C from K tiles

The kernel stages A/B values, multiplies register fragments and accumulates one C tile for two example K-tile iterations.

Accumulate C from K tilesThe kernel stages A/B values, multiplies register fragments and accumulates one C tile for two example K-tile iterations. A tile rows × reduction B tile reduction × columns C tile rows × columns × = Stage A/B use loaded values again Fragment MMA multiply + accumulate Partial C register accumulation next K tile
Figure 2. Staged A/B values supply data for an output tile through shared memory and register fragments. Partial C accumulates from K tiles. The 3×3 grids and two iterations show structure, not hardware tile dimensions or measured traces. The author made this explanatory diagram. Open full-size SVG