Block-Based Compression and Corresponding Hardware Circuits for Sparse Activations.

Area of Science:

• Computer Science
• Electrical Engineering
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) require significant memory and power, motivating research into reducing computational overhead.
• Exploiting sparsity in activation values is crucial for optimizing CNN accelerator performance.
• Existing methods primarily focus on sparsity, often overlooking the inherent similarity of feature maps within CNN layers.

Purpose of the Study:

• To propose a novel block-based compression approach for CNN activation values.
• To exploit both sparsity and similarity of activation values to minimize data volume.
• To reduce memory traffic and power consumption in CNN accelerators.

Main Methods:

• Developed a block-based compression strategy based on observations of activation value density and inter-channel similarity.
• Designed an encoder to convert output activations into a compressed format.
• Implemented a decoder and an indexing module to facilitate efficient computation with compressed data.

Main Results:

• The proposed approach effectively utilizes both sparsity and similarity of activation values.
• Benchmark data demonstrate substantial reductions in memory traffic.
• Consistent power consumption savings were observed compared to previous methods.

Conclusions:

• The block-based compression method offers a significant improvement in CNN accelerator efficiency.
• Exploiting feature map similarity alongside sparsity provides a more comprehensive optimization strategy.
• The designed encoder, decoder, and indexing modules effectively support the compression approach for practical applications.