Prefetches

Prefetches refer to the act of fetching data or instructions before they are actually needed, in order to hide memory latency and improve performance. They are a common feature in computer architecture, including CPUs and GPUs, and can also occur in web environments where browsers fetch resources ahead of user demand. Prefetching can be automatic, managed by hardware, or explicit, requested by software through compiler directives or intrinsics.

In computing, hardware prefetchers monitor the memory access patterns of a running program and attempt to predict future accesses to bring data into closer cache levels ahead of time. Common strategies include stride-based prediction, where repeated distances between accesses are detected, and more complex pattern recognition for irregular workloads. Software prefetching provides prefetch instructions or hints to the compiler, allowing programmers to specify addresses believed to be used soon. The effectiveness of prefetching depends on cache architecture, prefetch distance, and the regularity of memory access patterns.

Benefits of prefetching include reduced cache miss rates, fewer stall cycles, and improved instruction throughput. Drawbacks include cache pollution when predictions are incorrect, wasted bandwidth, and increased energy consumption. If prefetches arrive too late, they fail to hide latency; if too early, they may evict useful data. Therefore, prefetching works best for workloads with predictable access patterns and ample memory bandwidth.

In web contexts, prefetching involves speculatively fetching resources such as DNS entries or HTTP objects to accelerate page loads. This technique must balance performance gains against privacy, bandwidth, and security considerations.