Attention is memory-bound on long sequences. Tiled kernels fuse softmax with matmul so SRAM holds working blocks instead of materializing the full N×N matrix.
AI Mentor
Ask how anything on this page works…
↑↓ to navigate · Enter to jump · Esc to close
Your learning hub
DAY STREAK0
SECTIONS0
QUIZ BEST—
Notes for this section:—
Your progress is saved only in this browser. To continue on another device, export it to a file and import it there.
Your browser is blocking storage, so progress will be lost when you close this page. Use "Export my progress" to keep a copy.
How AI Works