EnforceSNN: Enabling resilient and energy-efficient spiking neural network inference considering approximate DRAMs...

Area of Science:

• Computer Engineering
• Artificial Intelligence
• Hardware Acceleration

Background:

• Spiking Neural Networks (SNNs) offer high accuracy and low power via bio-plausible computation.
• DRAM memory access is a major energy bottleneck in SNN processing.
• Existing SNN optimizations do not address DRAM energy-per-access.

Purpose of the Study:

• To develop a framework (EnforceSNN) for energy-efficient and resilient SNN inference using reduced-voltage DRAM.
• To enable SNNs in embedded systems with improved energy efficiency and performance.
• To mitigate potential errors introduced by approximate DRAM operation.

Main Methods:

• Quantized weights to lower DRAM access energy.
• Efficient DRAM mapping policy to minimize energy-per-access.
• Analysis of SNN error tolerance to varying bit error rates (BER).
• Fault-aware training (FAT) to enhance SNN resilience to DRAM errors.
• Algorithm for selecting SNN models balancing accuracy, memory, and energy.

Main Results:

• EnforceSNN maintains SNN accuracy for BER up to 10^-3.
• Achieved up to 84.9% DRAM energy savings.
• Realized up to 4.1x speed-up in DRAM data throughput.
• Demonstrated effectiveness across various network sizes.

Conclusions:

• EnforceSNN provides a viable solution for energy-efficient SNN inference on embedded systems.
• Reduced-voltage DRAM can be utilized effectively for SNNs with appropriate error mitigation.
• The framework significantly improves energy efficiency and performance without compromising accuracy.