LRMP: Layer Replication with Mixed Precision for spatial in-memory DNN accelerators.

Area of Science:

• Computer Engineering
• Artificial Intelligence
• Hardware Acceleration

Background:

• Deep Neural Networks (DNNs) face increasing computational demands, driving research into efficient hardware solutions.
• In-memory computing (IMC) utilizing non-volatile memories (NVMs) offers a promising avenue for accelerating DNNs through spatial parallelism.
• Existing NVM-based IMC accelerators struggle with non-uniform layer processing times and area constraints, limiting DNN performance.

Purpose of the Study:

• To develop a novel method, LRMP, for enhancing DNN performance on area-constrained NVM-based IMC accelerators.
• To address the challenges of non-uniform layer processing times and high area requirements in IMC architectures.
• To optimize DNN mapping by jointly considering layer replication and mixed-precision quantization.

Main Methods:

• LRMP employs a hybrid approach combining reinforcement learning and mixed integer linear programming.
• The method intelligently searches the design space of layer replication and mixed-precision quantization.
• A hardware-aware model guides the search, closely reflecting the target IMC accelerator architecture.

Main Results:

• LRMP demonstrates significant performance gains across five DNN benchmarks.
• Achieved 2.6-9.3x reduction in latency and 8-18x improvement in throughput.
• Maintained high accuracy with minimal degradation (<1%).

Conclusions:

• LRMP effectively optimizes DNN deployment on area-constrained NVM-based IMC accelerators.
• The joint application of layer replication and mixed precision quantization is crucial for performance enhancement.
• This method offers a practical solution for accelerating DNNs in resource-limited hardware environments.