Research

1 min read

Attention is 8% of the FLOPs and 41% of the wall clock

A layer-by-layer roofline of five frontier MoE models across four accelerators.

Pavle Padjin, Vladimir Zeljkovic

FLOP counts are the wrong unit for planning an inference deployment, and attention is the clearest example: a small share of the arithmetic, a large share of the time.

The measurement

Five frontier MoE models, four accelerators, one roofline per layer. We report arithmetic intensity, achieved bandwidth and achieved FLOP/s side by side, so each layer lands somewhere legible on the roof.

What it implies

If attention is memory-bound at your sequence length, buying arithmetic throughput buys you nothing. The decision that matters is layout and KV movement, and it is invisible in a FLOP budget.

Be there when
the next world opens.

Book a call