一句话：提出基于脉冲神经网络与STDP的Pathfinder预取器，以低开销实现在线实时学习并准确预测访存模式。
数据预取对高性能处理器至关重要，但已有基于深度神经网络的预取器训练和推理开销高，无法实时训练，只能依赖事先用大规模基准训练的大模型，对未见访问模式泛化差。本文提出 Pathfinder，用脉冲神经网络（SNN）和 STDP 局部学习算法在线学习访存增量模式，在推理同时进行低成本训练，从而快速适应当前访问流。作者在多种基准上评估，Pathfinder 与最先进预取器性能相当；在 12nm 工艺下功耗仅 0.5 W、面积 0.23 mm²。结果表明神经启发预取可兼具实用性与高性能，但 SNN/STDP 预取算法仍需更多创新以充分释放潜力。
Data prefetching is vital in high-performance processors and a large body of research has introduced a number of different approaches for accurate prefetching: stride detection, address correlating prefetchers, delta pattern detection, irregular pattern detection, etc. Most recently, a few works have leveraged advances in machine learning and deep neural networks to design prefetchers. These neural-inspired prefetchers observe data access patterns and develop a trained model that can then make accurate predictions for future accesses. A significant impediment to the success of these prefetchers is their high implementation cost, for both inference and training. These models cannot be trained in real-time, i.e., they have to be trained beforehand with a large benchmark suite. This results in a large model (that increases the overhead for inference), and the model can only successfully predict patterns that are similar to patterns in the training set. In this work, we explore the potential of using spiking neural networks to learn and predict data access patterns, and specifically address deltas. While prior work has leaned on the recent success of trained artificial neural networks, we hypothesize that spiking neural networks are a better fit for real-time data prefetching. Spiking neural networks rely on the STDP algorithm to learn while performing inference - it is a low-cost and local learning algorithm that can quickly observe and react to the current stream of accesses. It is therefore possible to achieve high accuracy on previously unseen access patterns with a relatively small spiking neural network. This paper makes the case that spiking neurons and STDP offer a complexity-effective approach to leverage machine learning for data prefetching. We show that the proposed Pathfinder prefetcher is competitive with other state-of-the-art prefetchers on a range of benchmarks. Pathfinder can be implemented at 0.5 W and an area footprint of only 0.23 mm2 (12 nm technology). This work shows that neural-inspired prefetching can be both practical and capable of high performance, but more innovations in SNN/STDP prefetch algorithms are required to fully maximize their potential.
@inproceedings{JiaMGSB24,
title = {PATHFINDER: Practical Real-Time Learning for Data Prefetching},
author = {Lin Jia and James Patrick Mcmahon and Sumanth Gudaparthi and Shreyas Singh and Rajeev Balasubramonian},
booktitle = {ASPLOS},
year = {2024},
doi = {10.1145/3620666.3651332}
}