Modules B05–B06 · concept player

Same weights, two machines

One projection, one formula: intensity = Td / (d + 2T). Slide tokens-in-flight and watch decode at batch 1 sit on the floor, then climb past the machine's ridge point into prefill territory as batching or sequence length grows.

Hidden size d
Tokens in flight, T (batch × sequence): 1
1.0FLOP/byte, this T
Decodewhich machine this looks like
281T where intensity = ridge (247)

Intensity as T grows

Exact arithmetic on the module's formula, not a measurement. At d=4096 this reproduces B05's own worked points exactly: 1.0 at T=1, roughly 247 at T=281 (the machine's cited ridge point on MI300X), 1024 at T=2048. B05 section 2 derives the formula; decode at batch 1 uses about 0.4% of the machine's arithmetic because the cost there is entirely the weight read, not the multiply.