Grow the batch until parallelism stops buying progress.

Gradient noise creates a critical batch size: below it, larger batches improve update quality; above it, extra examples mostly duplicate information. Explore that bend directly.

Batch efficiency curve

Update efficiency0%
Critical ratio0x
Step time0 ms
Suggested LR multiplier0x

One curve connects statistics and systems.

Small batches are noise-limited.

Adding examples reduces gradient variance and makes each optimizer step more reliable.

Critical batch marks the bend.

Near the gradient noise scale, statistical efficiency begins to saturate.

Large batches become systems bets.

More workers can shorten wall time only if communication and stragglers stay controlled.

Primary reading

An Empirical Model of Large-Batch TrainingAccurate, Large Minibatch SGDDo not decay the learning rate, increase the batch size