How busy can one chip stay?

A chip (a GPU) does arithmetic very fast but waits for numbers to arrive from its memory. Drag the slider: the more work it gets from each delivery, the less it waits.

Chip busy
Change the chip
Work per byte fetched (FLOP/byte)
Work per byte needed to stay busy (the ridge)
Speed reached (TFLOP/s)
FLOP = one add or multiply. TFLOP/s = trillions per second. Top speed in TFLOP/s, memory speed in GB/s. Numbers are 2 bytes each (FP16). The A100 button sets the data-sheet values of the 80 GB SXM card: 312 TFLOP/s and 2,039 GB/s. Chatbot mode multiplies a 4096 by 4096 weight grid by M words (one per user, or a long prompt); tile mode multiplies two 4096 by 4096 grids with tiles b wide. Both axes are logarithmic. Grey dots mark the 13 slider stops.