Matmul is the force

C = AB. Hover a cell. ~90% of GPU arithmetic.

play

A m×k

B k×n

C m×n

hover a cell of C to see its row × column · click it to accumulate

rows of A · m 3

shared dim · k 4

cols of B · n 3

output cells9

MACs / cell4

total MACs36

FLOPs ≈ 2mnk72

Hover a C cell. Click it. That multiply is ~90% of GPU work.

C_{ij} = Σ_k A_{ik} B_{kj}

FLOPs = 2mnk. Intensity = FLOPs / bytes. That ratio picks the roofline.