Why it matters

Perf analysis picks the right optimization. Wrong target = wasted eng weeks.

Advertisement

The architecture

Decompose: prefill, decode, transport.

Attribute to compute vs memory vs comms.

Guide next optimization.

LLM perf analysisDecomposeprefill, decode, netAttributecompute vs bandwidthGuidenext optimizationRoofline model attributes; guide investment where impact is highest
LLM perf analysis.
Advertisement

How it works end to end

Roofline model.

Prefill compute-bound; decode memory-bound.

Comms in multi-GPU.