A Digital Processing in Memory Architecture Using TCAM for Rapid Learning and Inference Based on a Spike Location Dependent Plasticity

Seong-Min Kim, Kyeong Min Kim, Ji Hoon Choi, Seung Man Kang, Min Kyung Bang, So Hee Park, Eon Gyeong Lee, Seong Bae Park, Choong Seon Hong, Sang Hoon HongIEEE Access 2023 · DOI

Abstract

The proliferation of Artificial Intelligence of Things (AIoT) devices necessitates efficient "edge learning" capabilities. Traditional Deep Neural Networks (DNNs) are often too computationally expensive for battery-powered devices due to iterative backpropagation. This paper describes a purely digital processing-in-memory (DPIM) architecture that leverages Ternary Content Addressable Memory (TCAM) to achieve high-speed learning and inference. By utilizing a novel Spike Location Dependent Plasticity (SLDP) rule, the system treats learning as a frequency-based sorting problem rather than a weight-optimization task. Experimental results on the MNIST dataset demonstrate 94.58% accuracy, millisecond-scale training, and sub-4mW power consumption.

Introduction

As AI moves to the edge, the demand for local learning—the ability of a device to adapt to new data without cloud connectivity—has increased. However, the energy cost of traditional Multiply-Accumulate (MAC) operations and backpropagation remains a significant barrier. While Spiking Neural Networks (SNNs) offer a bio-inspired, low-power alternative, they often face challenges in scalability and hardware implementation.

Processing-in-Memory (PIM) has emerged as a solution to the "memory wall" by performing computations directly where data is stored. While many PIM designs are analog, they are susceptible to environmental noise and process variations. This research proposes a robust, fully digital PIM architecture using TCAMs to enable rapid, low-power edge learning.

The SE-SFNN Architecture

The core of the proposed system is the Stride Edge-Detection Search Frequency Neural Network (SE-SFNN). This architecture is designed to extract features and classify them using a two-layer digital approach. The first layer utilizes digital filtering kernels (horizontal, vertical, and diagonal) to identify edges. The second layer performs classification based on the frequency of these patterns.

A critical component of this architecture is the Ternary Content Addressable Memory (TCAM). Unlike standard binary memory, TCAMs support a "don't care" state (X), allowing for flexible pattern matching.

TCAM operation with 'don't care' bits (X) allowing for flexible pattern matching.

By using TCAMs, the system can ignore irrelevant pixels within a 3x3 feature block, making the search process robust against minor input variations and reducing the computational overhead of exact matching.

Spike Location Dependent Plasticity (SLDP)

The learning mechanism, Spike Location Dependent Plasticity (SLDP), is a simplified digital adaptation of biological Spike Timing Dependent Plasticity (STDP). Instead of tracking precise spike timings, SLDP focuses on the location or rank of patterns within the TCAM.

The architecture separates the data flow into an inference path and a training path. During training, the system monitors the frequency of specific 3x3 pixel patterns using a set of counters.

The DPIM architecture showing the inference and training paths, including the counter-based ranking system.

Frequent patterns are promoted to higher addresses in the TCAM. This "Minor to Major Promotion" (M2MP) scheme acts as a temporal cache, keeping only the most relevant visual features. This approach effectively transforms learning into a sorting problem, eliminating the need for power-hungry multipliers and complex weight updates.

Results and Performance

The architecture was validated using the MNIST handwritten digit dataset. The system achieved a peak accuracy of 95.69%, which slightly adjusted to 94.58% when using the memory-optimized M2MP scheme.

The performance metrics are particularly notable for edge applications:

  • Speed: Training on a set of 6,000 images was completed in just 6.3 milliseconds at a 100 MHz clock frequency.
  • Throughput: The hardware supports an inference rate of over 5.8 million images per second.
  • Power: The average power dissipation is less than 4 mW.
  • Efficiency: The M2MP scheme reduced the required TCAM size by 84.4% and decreased the clock cycles needed for learning by over 97%.

Discussion and Limitations

While the SE-SFNN architecture excels at grayscale digit recognition, scaling to more complex datasets like CIFAR-10 presents challenges. The current ternary representation is relatively coarse, and future work will likely require higher-resolution quantization to handle color and complex textures. Additionally, while the system is robust to digital noise, its accuracy on the noisy N-MNIST dataset (89.94%) suggests that further refinement of the "stride and neighbor" logic is needed for high-interference environments.

Conclusion

This digital PIM architecture represents a significant step toward viable edge learning. By replacing complex weight updates with a frequency-based ranking system in TCAM, the researchers have created a hardware-efficient model that is fast, low-power, and environmentally robust. This approach paves the way for a new generation of AIoT devices capable of real-time local adaptation.

Summary updated

BibTeX
@article{kim2023digital,
  author = {Seong-Min Kim and Kyeong Min Kim and Ji Hoon Choi and Seung Man Kang and Min Kyung Bang and So Hee Park and Eon Gyeong Lee and Seong Bae Park and Choong Seon Hong and Sang Hoon Hong},
  title = {{A Digital Processing in Memory Architecture Using TCAM for Rapid Learning and Inference Based on a Spike Location Dependent Plasticity}},
  journal = {IEEE Access},
  year = {2023},
  doi = {10.1109/access.2023.3234323},
}
Figure