1 A Survey of Machine Learning Applied to Computer Architecture Design Drew D. Penney, and Lizhong Chen∗, Senior Member, IEEE Abstract—Machine learning has enabled significant benefits in diverse fields, but, with a few exceptions, has had limited impact on computer architecture. Recent work, however, has explored broader applicability for design, optimization, and simulation. Notably, machine learning based strategies often surpass prior state-of-the-art analytical, heuristic, and human-expert approaches. This paper reviews machine learning applied system-wide to simulation and run-time optimization, and in many individual components, including memory systems, branch predictors, networks-on-chip, and GPUs. The paper further analyzes current practice to highlight useful design strategies and identify areas for future work, based on optimized implementation strategies, opportune extensions to existing work, and ambitious long term possibilities. Taken together, these strategies and techniques present a promising future for increasingly automated architectural design. F 1 INTRODUCTION In the past decade, machine learning (ML) has rapidly become a revolutionary factor in many fields, ranging from commercial applications, as in self-driving cars, to medical applications, improving disease screening and diagnosis. In each of these applications, an ML model is trained to make predictions or decisions without explicit programming by discovering embedded patterns or relationships in the data. Notably, ML models can perform well in tasks/applications where relationships are too complex to model using analyt- ical methods. These powerful learning capabilities continue to enable rapid developments in diverse fields. Concur- rently, the exponential growth predicted by Moore’s law has slowed, putting increasing burden on architects to supplant Moore’s law with architectural advances. These opposing Fig. 1. Publications on machine learning applied to architecture (for trends suggest opportunities for a paradigm shift in which works examined in Section3) computer architecture enables ML and, simultaneously, ML improves computer architecture, closing a positive-feedback tinue to provide breakthroughs as promising applications loop with vast potential for both fields. are explored. Traditionally, the relationship between computer archi- The paper is organized as follows. Section 2 provides tecture and ML has been relatively imbalanced, focusing background on ML and existing models to build intuition on on architectural optimizations to accelerate ML algorithms. ML applicability to architectural issues. Section 3 presents In fact, the recent resurgence in AI research is, at least existing work on ML applied to architecture. Section 4 partly, attributed to improved processing capabilities. These then compares and contrasts implementation strategies in improvements are enhanced by hardware optimizations existing work to highlight significant design considerations. exploiting available parallelism, data reuse, sparsity, etc. in Section 5 identifies possible improvements and extensions existing ML algorithms. In contrast, there has been relatively to existing work as well as promising, new applications for limited work applying ML to improve architectural design, future work. Finally, Section 6 concludes. arXiv:1909.12373v1 [cs.AR] 26 Sep 2019 with branch prediction being one of a few mainstream examples. This nascent work, although limited, presents an 2 BACKGROUND auspicious approach for architectural design. 2.1 Fundamental Applicability This paper presents an overview of ML applied to archi- tectural design and analysis. As illustrated in Figure1, this Machine learning has been rapidly adopted in many fields field has grown significantly in both success and popularity, as an alternative approach for a diverse range of prob- particularly in the past few years. These works establish lems. This fundamental applicability stems from the pow- the broad applicability and future potential of ML-enabled erful relationship learning capabilities of ML algorithms. architectural design; existing ML-based approaches, ranging Specifically, ML models leverage a generic framework in from DVFS with simple classification trees to design space which these models learn from examples, rather than ex- exploration via deep reinforcement learning, have already plicit programming, enabling application in many tasks, surpassed their respective state-of-the-art human expert and including those too difficult to representing using standard heuristic based designs. ML-based design will likely con- programming methods. Furthermore, using this generic framework, there may be many possible approaches for ∗ any given problem. For example, in the case of predicting Corresponding author. Email: [email protected] IPC for a processor, one can experiment with a simple The authors are with Oregon State University, Corvallis, OR 97331 Copyright 2019 by Drew D. Penney and Lizhong Chen linear regression model, which learns a linear relationship All Rights Reserved between features (such as core frequency and cache size) 2 and the instructions-per-cycle (IPC). This approach may sums. Additional variants such as convolutional neural net- work well or it may work poorly. In the case it works works (CNNs) incorporate convolution operations between poorly, one can try different features, non-linear feature some layers to capture spatial locality while recurrent neural combinations (such as core frequency times cache size), networks re-use the previous output to learn sequences and or a different model entirely, with another common choice long-term patterns. All these supervised learning models being an artificial neural network (ANN). This diversity in can be used in both classification and regression tasks, possible approaches enables adjustment of models, model although there are some distinct high-level differences. Vari- parameters, and training features to match the task at hand. ants of SVMs and neural networks tend to perform better for high-dimension and continuous features and also when features may be nonlinear [1]. These models, however, tend 2.2 Learning Approaches & Models to require more data compared to Bayesian networks and The learning approach and the model are both fundamental decision trees. considerations in applying machine learning to any prob- Unsupervised learning: Unsupervised learning uses just lem. In general, there are four main categories of learn- input data to extract information without human effort. ing approaches: supervised learning, unsupervised learn- These models can therefore be useful, for example, in reduc- ing, semi-supervised learning, and reinforcement learning. ing data dimensionality by finding appropriate alternative These approaches can be differentiated by what data is used representations or clustering data into classes that may not and how that data is used to facilitate learning. Similarly, be obvious for humans. many appropriate models may exist for a given problem, Thus far, the primary two unsupervised learning models thus enabling significant diversity in application based applied to architecture are principal components analysis on the learning approach, hardware resources, available (PCA) and k-means clustering. PCA provides a method data, etc. In the following, we introduce these learning to extract significant information from a dataset by deter- approaches and several significant models for each learning mining linear feature combinations with high variance [4]. approach, focusing on approaches with proven applicability. As such, PCA can be applied as an initial step towards Implementation details are considered later in Section4. building a model with reduced dimensionality, a highly Supervised learning: In supervised learning, the model desirable feature in most applications, albeit at the cost of is trained using input features and output targets, with the interpretability (discussed in Section4). K-means clustering result being a model that can predict the output for new, is instead used to identify groups of data with similar fea- unseen inputs. Common supervised learning applications tures. These groups may be further processed to generalize include regression (predicting a value such as processor behavior or simplify representations for large datasets. IPC) and classification (predicting a label such as the op- Semi-supervised learning: Semi-supervised learning timal core configuration for application execution). Feature represents a mix of supervised and unsupervised methods, selection, discussed in Section 2.3, is particularly important with some paired input/output data, and some unpaired in- in these applications as the model must learn to predict put data. Using this approach, learning can take advantage solely based on feature values. of limited labeled data and potentially significant unlabeled Supervised learning models can be generalized into data. We note that this approach has, thus far, not yet four categories: decision trees, Bayesian networks, support found application in architecture. Nevertheless, one work vector machines (SVMs), and artificial neural networks [1]. on circuits analysis [5] presents a possible strategy that Decision trees use a tree structure where each node repre- could be adapted in future work. sents a feature and branches represent a value (or range of Reinforcement Learning: In reinforcement learning, an values) for that feature. Inputs are therefore classified by agent is sequentially provided with