Why it matters

Throughput math shapes ops economics. Understanding shapes design.

Advertisement

The architecture

Per GPU: HBM_BW / bytes_per_token / bytes_per_param.

Batch amortizes.

Throughput scalingPer-stream tokens/secsingle batchBatch amortizationshared weight readsTotal throughputbatch * per-streamDiminishing returns: batch beyond ~64 gains little; KV cache limits practical batch
Throughput.
Advertisement

How it works end to end

Continuous batching scales throughput.

KV cache limits practical batch.

Speculative decoding boosts single-stream.