Digit-Serial DA-Based Fixed-Point RNNs: A Unified Approach for Enhancing Architectural Efficiency

Area of Science:

• Computer Engineering
• Artificial Intelligence
• Embedded Systems

Background:

• Integrating advanced artificial intelligence (AI) neural network models into resource-constrained embedded and mobile systems is a critical challenge.
• Developing compact, energy-efficient neural network models in silicon is essential for optimizing performance in these systems.
• Long short-term memory (LSTM) recurrent neural networks (RNNs) are powerful but computationally intensive, requiring architectural enhancements for efficiency.

Purpose of the Study:

• To present a unified approach for enhancing the architectural efficiency of LSTM RNNs.
• To introduce two novel structures (Structure I and Structure II) for efficient computation within LSTM RNNs.
• To validate the proposed structures for speech recognition tasks using fixed-point quantized models.

Main Methods:

• Developed two new LSTM RNN structures (I and II) utilizing the two's complement (TC) digit-serial distributed arithmetic (DSDA) technique.
• Formulated block-circulant matrix-vector multiplications (MVMs) and element-wise multiplications (EWMs) using TC DSDA.
• Introduced a new partial-product generator (PPG) and partial-product selector (PPS) compatible with signed and unsigned digits, and employed a fixed-point (FxP) training procedure for quantized LSTM RNNs.

Main Results:

• Structure I uses a non-pipelined MVM with a shift-accumulate unit (SAU), while Structure II employs a depth-pipelined MVM with a shift-to-add tree (SAT).
• Post-synthesis results on 28-nm FDSOI technology show significant improvements: Structure I offers 39.87% (area) and 30.95% (energy) efficiency gains.
• Structure II demonstrates even greater gains, achieving 95.63% (area) and 91.18% (energy) efficiency improvements for a $128 \times 128$ model size compared to previous works.

Conclusions:

• The proposed TC DSDA-based structures provide substantial enhancements in architectural efficiency for LSTM RNNs.
• These structures are highly effective for implementing quantized LSTM RNNs in silicon, particularly for applications like speech recognition.
• The novel PPG and PPS components contribute to the overall efficiency and flexibility of the proposed designs.