recompute

Recompute is the act of computing a result again, rather than using a previously stored value. In computing and data processing, recomputing may occur when inputs have changed, when a cached intermediate result is unavailable, or when a pipeline is designed to favor recomputation over storage. Recomputing can help ensure correctness and determinism if the underlying algorithm is deterministic.

Applications include data pipelines, simulations, rendering, and large-scale analytics where keeping all intermediate results would be expensive in memory. In such contexts, systems may discard temporary data and recompute as needed to reduce storage footprints or to simplify fault recovery. Recompute is also a consideration in streaming and incremental updates, where recomputing from base data can avoid stale caches.

In machine learning and automatic differentiation, recomputation refers to a memory-saving technique often called checkpointing or activation recomputation. Rather than storing all intermediate activations during a forward pass, a model may selectively discard them and recompute them during backpropagation. This reduces peak memory usage at the cost of additional forward and backward compute time. Frameworks such as PyTorch and TensorFlow offer tools or patterns to implement gradient checkpointing.

Trade-offs include computational expense, latency, energy consumption, and determinism. Recompute is advantageous when memory is the limiting resource, when storage costs are high, or when fault tolerance requires recomputing from base inputs. It is less suitable when recomputation would noticeably slow down real-time or interactive workloads or when non-determinism is present.

See also gradient checkpointing.