MAx-DNN: Multi-Level Arithmetic Approximation for Energy-Efficient DNN Hardware Accelerators

2025-06-26

AI Summary

Addressing the challenge of balancing energy efficiency and accuracy in low-power DNN hardware acceleration, this paper proposes a three-level collaborative approximation method—operating at the layer, filter, and convolution kernel granularities—and introduces the first dynamic, multi-granularity deployment of ROU-P approximate multipliers, precisely aligned with the error-resilience characteristics of individual DNN layers. Leveraging quantization-aware approximate modeling and fine-grained error distribution optimization, the approach overcomes the limitations of conventional coarse-grained, single-level approximation. Experimental evaluation on ResNet-8/CIFAR-10 demonstrates a 54% energy-efficiency improvement over the baseline quantized model (with ≤4% accuracy loss) and achieves twice the energy efficiency of state-of-the-art DNN approximation methods while maintaining higher accuracy. This work establishes a novel co-design paradigm integrating approximate computing with neural network architecture optimization.

Technology Category

• Machine Learning: Hardware-aware ML
• Computer Vision: Learning & Optimization for CV
• Search and Optimization: Learning to Search

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Systems and Infrastructure for Web, Mobile and WoT: Energy management for devices in mobile Web and WoT environments

Abstract

Nowadays, the rapid growth of Deep Neural Network (DNN) architectures has established them as the defacto approach for providing advanced Machine Learning tasks with excellent accuracy. Targeting low-power DNN computing, this paper examines the interplay of fine-grained error resilience of DNN workloads in collaboration with hardware approximation techniques, to achieve higher levels of energy efficiency. Utilizing the state-of-the-art ROUP approximate multipliers, we systematically explore their fine-grained distribution across the network according to our layer-, filter-, and kernel-level approaches, and examine their impact on accuracy and energy. We use the ResNet-8 model on the CIFAR-10 dataset to evaluate our approximations. The proposed solution delivers up to 54% energy gains in exchange for up to 4% accuracy loss, compared to the baseline quantized model, while it provides 2x energy gains with better accuracy versus the state-of-the-art DNN approximations.

Problem

• Enhancing energy efficiency in DNN hardware accelerators
• Balancing accuracy loss with energy savings in DNNs
• Optimizing approximate multipliers for low-power DNN computing

Innovation

• Multi-level arithmetic approximation for DNN accelerators
• Fine-grained error resilience with hardware approximation
• Layer-filter-kernel level ROUP multiplier distribution