Why it matters

BW-bound ops waste GPU. Understanding shapes optimization.

Advertisement

The architecture

Compare: FLOPs / bytes = arithmetic intensity.

Below balance: BW-bound.

BW-bound opsArithmetic intensityFLOPs / byteGPU balanceTFLOPs / TB/sBW-boundif below balanceH100: 989 TFLOPS bf16 / 3.35 TB/s = ~300 FLOP/byte balance
BW-bound.
Advertisement

How it works end to end

H100 balance: ~300 FLOP/byte for bf16 tensor.

Attention softmax: BW-bound.

LayerNorm, GELU: BW-bound.

Fuse into matmul for compute-bound.