Reproduce “Full-Stack Optimization for CAM-Only DNN Inference” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

João Paulo C. de Lima, Asif Ali Khan, Luigi Carro, Jeronimo Castrillon

An integrated hardware-software approach leverages Racetrack Memory (RTM)-based Content-Addressable Memory (CAM) associative processors and a specialized compiler to optimize Deep Neural Network inference. This system achieves a 7.5x energy efficiency gain for ResNet-18 on ImageNet with 4-bit activations and reduces activation data movement to 3% of total energy, maintaining accuracy comparable to software.

Key insights

  • A cohesive full-stack hardware-software co-design is essential for effectively overcoming the memory wall in DNN inference and realizing the full potential of Computing-in-Memory.
  • Ternary Weight Networks (TWNs) combined with moderately quantized (4-bit) activations can maintain high accuracy in DNNs while enabling highly energy-efficient bit-wise operations in CAM-based architectures.
  • Strategically leveraging the unique physical characteristics of emerging memories, such as RTM's sequential access for bit-serial processing, allows for the creation of more dense and efficient computational units.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.