sys·to·le /ˈsɪstəli/ — the contraction of the heart that pumps blood outward. A TPU multiplies matrices the same way: data pulsing through a grid of multiply–accumulate cells, one beat at a time. This is W × I, live.
etched into the die before the run
skewed queue, one column per beat
drips from the bottom row
what gets pumped through the die
The amber band is the computation wavefront — every cell inside it fires one multiply–accumulate per beat, deepening as sums gather downward. Blue texture inside the band is the input data itself, streaming right. Drag the scrubber — any beat is computable instantly.
An n×n product needs n³ multiply–accumulates. The array does n² of them every beat, finishing in 3n−2 beats — that's the whole trick.
Weights load transposed — cell (r,c) holds W[c][r] — so each
column of the die accumulates one row of W×I as sums descend. The staggered
queue makes every activation meet its partial sum at exactly the right beat; an n×n product
drains in 3n−2 beats. Everything runs in your browser — micro sizes step a
cycle-accurate simulator, macro sizes render the same dataflow analytically.
space play → step r reset