the.ai

Systems / Training

verified

Mixed-Precision Training

Most of training does not need full precision. Doing the arithmetic in half precision roughly doubles throughput and halves activation memory, while keeping a full-precision copy of the weights so that many tiny updates still accumulate instead of rounding away to nothing.

Viz primitive · budget-splitactivations = 8

activations holds 50% of the budget; rest holds the remaining 50%.

Memory held by half-precision activations against the full-precision state beside them. Drag the activation size to see which half of the budget precision actually saves.

8

Reviewed by opendroid · 2026-08-04