Onur Mutlu Lectures

Digital Design & Computer Architecture - Lecture 25: Prefetching (Spring 2022)

The penultimate lecture of the Spring 2022 Digital Design & Computer Architecture series focuses on prefetching – the proactive loading of data into cache or registers before it is demanded by the processor. The instructor emphasizes that prefetching is one of the most impactful techniques for alleviating memory‑hierarchy bottlenecks, especially as systems adopt deeper, heterogeneous memory stacks. Key insights include how prefetching can cut both cache‑miss rates and miss‑latency by bringing blocks into the appropriate cache level ahead of time. The design space spans from L1 to main memory and even remote nodes, raising coordination challenges across levels. The lecture also covers software‑driven prefetching (e.g., browsers preloading likely links) and hardware mechanisms, noting that aggressive strategies may introduce inter‑core interference and bandwidth pressure. The professor highlights concrete examples: a web browser predicting user clicks, distributed prefetches from remote memory nodes, and the subtle interplay with cache‑coherence protocols such as MESI. He warns that a prefetch triggered by one core can evict useful data for another, underscoring the need for intelligent prediction and throttling. For system architects, the takeaway is clear: effective prefetching can unlock significant performance gains, but only if designers carefully balance prediction accuracy, placement, and coherence overhead. As multicore and emerging memory technologies (e.g., PCM, FeRAM) become mainstream, refined prefetch strategies will be essential to sustain scaling and latency targets.