Mixed-precision quantization techniques for energyefficient DNN inference

Key Points

• Results indicate that mixed precision quantization significantly reduces model bitwidth assignments, enhancing computational efficiency.
• Two quantization-aware training methods were implemented, showcasing a potential to maintain accuracy while lowering resource demands.
• The analysis showed the ability to deploy neural networks with reduced bitwidth, affirming their feasibility in practical applications.
• Implications suggest that such techniques can lead to energy-efficient inference, which is crucial for large-scale deployment of deep learning models.

Abstract