TRAM: Joint Optimization of Approximate Multipliers and AI Models for Low-Power Accelerators
Researchers have introduced TRAM, a novel framework designed to reduce power consumption in AI accelerators by jointly optimizing approximate multiplier (AxM) structures and AI model parameters. Unlike previous methods that design hardware components separately from software training, TRAM integrates these processes to minimize energy usage while maintaining high accuracy. Multipliers are identified as significant power consumers in AI models, making them a primary target for optimization through approximate computing. Experimental results demonstrate the effectiveness of this approach, showing that TRAM achieves up to a 25.05% reduction in AxM power consumption on Convolutional Neural Networks (CNNs) using the CIFAR-10 dataset. Furthermore, it reduces power by up to 27.09% on Vision Transformers tested with ImageNet. This study highlights the potential of co-designing hardware and algorithms to enhance the energy efficiency of modern AI systems, addressing the growing demand for sustainable and low-power computing solutions in machine learning applications.
Wire timeline
TRAM: Joint Optimization of Approximate Multipliers and AI Models for Low-Power Accelerators
Researchers have introduced TRAM, a novel framework designed to reduce power consumption in AI accelerators by jointly optimizing approximate multiplier (AxM) structures and AI model parameters. Unlike previous methods that design hardware components separately from software training, TRAM integrates these processes to minimize energy usage while maintaining high accuracy. Multipliers are identified as significant power consumers in AI models, making them a primary target for optimization through approximate computing. Experimental results demonstrate the effectiveness of this approach, showing that TRAM achieves up to a 25.05% reduction in AxM power consumption on Convolutional Neural Networks (CNNs) using the CIFAR-10 dataset. Furthermore, it reduces power by up to 27.09% on Vision Transformers tested with ImageNet. This study highlights the potential of co-designing hardware and algorithms to enhance the energy efficiency of modern AI systems, addressing the growing demand for sustainable and low-power computing solutions in machine learning applications.
cs.AI updates on arXiv.org