A Survey of Machine Learning Applied to Computer Architecture Design

Drew D. Penney, Lizhong Chen

Abstract

As the complexity of modern hardware increases and Moore’s Law slows, traditional human-expert heuristics are becoming insufficient for optimal computer architecture design. This paper surveys the emerging field of machine learning (ML) applied to hardware design, simulation, and optimization. We categorize ML applications into offline design-space exploration and online runtime control, highlighting significant performance improvements across branch prediction, cache management, and network-on-chip (NoC) design. The survey identifies a paradigm shift toward data-driven hardware design, while addressing the challenges of implementation overhead and model interpretability.

Introduction

Traditionally, computer architecture has focused on creating specialized hardware to accelerate machine learning workloads. However, a "reverse" relationship is now gaining momentum: the application of machine learning to the design and optimization of the hardware itself. As modern systems-on-chip (SoCs) and GPUs grow in complexity, the design space becomes too vast for manual navigation using traditional analytical models.

Architects are increasingly turning to ML to replace rigid, human-designed heuristics with flexible, data-driven models. This transition is driven by the need for more automated and efficient design processes that can discover hidden patterns in architectural data. The research interest in this domain has seen an exponential rise, particularly in the last decade.

Publication trends showing the rapid growth of ML applications in computer architecture.

Taxonomy of Machine Learning Applications

The application of ML in computer architecture can be broadly categorized into two operational contexts:

1. Offline Applications

These involve using ML to accelerate design-time tasks. A primary use case is creating "surrogate models" for system simulation. Cycle-accurate simulators are computationally expensive; ML models can provide over 90% accuracy while being several orders of magnitude faster, allowing for much broader design space exploration (DSE).

2. Online Applications

These involve integrating ML models directly into the hardware logic to make real-time decisions. Examples include:

  • Branch Prediction: Using neural networks (like perceptrons) to predict instruction flow.
  • Memory Control: Using Reinforcement Learning (RL) to manage scheduling and maximize throughput.
  • Power Management: Implementing dynamic voltage and frequency scaling (DVFS) through learned policies.

Methodologies and Benchmarking

The field utilizes a diverse set of ML techniques tailored to specific architectural problems. Supervised learning is frequently used for performance (IPC) and power prediction, while Reinforcement Learning (RL)—specifically Q-learning and Deep Q-Networks (DQN)—is preferred for control problems like NoC routing. Unsupervised learning techniques like PCA help in reducing the dimensionality of complex design spaces.

To evaluate these models, researchers rely on standard industry benchmarks:

  • SPEC Benchmarks: For general-purpose CPU modeling.
  • DirectX Applications: For GPU performance analysis.
  • Synthetic Traffic: For evaluating Network-on-Chip (NoC) efficiency.
  • Simulated Datasets: Generated from tools like Gem5 or GPGPU-Sim to train surrogate models.

Key Results and Performance Gains

The integration of ML has yielded impressive quantitative improvements over traditional methods:

  • Memory Systems: RL-based memory controllers have demonstrated a 15% performance improvement over traditional scheduling heuristics.
  • Interconnects: ML-based power scaling in photonic NoCs reduced laser power consumption by 42% with negligible impact on throughput.
  • GPU Optimization: Frameworks like Stargazer achieved performance prediction errors of less than 3.8% across nearly a million design points using regression models.
  • Simulation Efficiency: ML models have successfully modeled Chip Multiprocessor (CMP) performance with only 4-5% error while requiring training on only 1% of the total design space.

Challenges and Future Directions

Despite its potential, several hurdles remain for widespread adoption:

  • Hardware Overhead: On-chip ML models must be lightweight. If a model consumes too much area or power, it may negate the performance benefits it provides.
  • Interpretability: The "black box" nature of deep learning makes hardware validation difficult. Understanding the rationale behind an ML-driven architectural decision is critical for reliability.
  • Generalization: There is a growing need for generalized tools—a "scikit-learn for architects"—to standardize ML application in hardware research.

Future research is expected to move toward hierarchical and multi-agent ML systems, where distributed learning agents cooperate across different chip components to optimize global system behavior.

Conclusion

The application of machine learning to computer architecture represents a fundamental shift from heuristic-based design to empirical, data-driven optimization. By acting as high-speed surrogates for simulation and adaptive controllers for runtime management, ML models are enabling the next generation of high-performance, energy-efficient hardware.

Summary updated

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.
BibTeX
@misc{penney2019survey,
  author = {Drew D. Penney and Lizhong Chen},
  title = {{A Survey of Machine Learning Applied to Computer Architecture Design}},
  year = {2019},
  eprint = {1909.12373},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1909.12373},
}
Figure