Flash-based content addressable memory with L2 distance for memory-augmented neural network.

Area of Science:

• Computer Science
• Electrical Engineering
• Artificial Intelligence

Background:

• Memory-Augmented Neural Networks (MANN) are crucial for lifelong on-device learning.
• Efficient explicit memory implementation is vital for MANN performance.
• Content Addressable Memory (CAM) offers in-memory computing for accelerated memory access.

Purpose of the Study:

• To propose a novel CAM cell with quadratic coding for enhanced MANN performance.
• To fabricate and evaluate a 1Mb Flash-based multi-bit CAM chip for Euclidean distance computation.
• To demonstrate the robustness and energy efficiency of the proposed CAM solution.

Main Methods:

• Design and fabrication of a 1Mb Flash-based multi-bit CAM chip utilizing a quadratic code CAM cell.
• Implementation of Euclidean (L2) distance computation within the CAM.
• Performance evaluation on the Omniglot dataset and comparison with ternary CAM and GPUs.
• Assessment of the chip's robustness under environmental stress (baking at 200°C for 10^5 s).

Main Results:

• Significant reductions in latency (5.3x) and energy (46.6x) compared to ternary CAM for MANN on Omniglot.
• Minimal recognition accuracy degradation (<1%) after prolonged high-temperature exposure, indicating robustness.
• Drastic performance improvements over GPUs: 471-fold reduction in latency and 1267-fold in energy for search operations.
• Successful fabrication of a 1Mb multi-bit CAM chip demonstrating quadratic code functionality.

Conclusions:

• The proposed CAM cell and chip offer a robust and highly energy-efficient solution for explicit memory in MANNs.
• This technology enables practical implementation of lifelong on-device machine intelligence.
• The CAM chip significantly advances the state-of-the-art in in-memory computing for AI applications.