Reconfigurable Multiplier Architecture for Error-Resilient RISC-V Cores
Researchers have introduced a novel runtime reconfigurable multiplier architecture integrated into the RISC-V core, specifically designed to enhance energy efficiency in neural network inference and edge AI applications. Addressing the critical challenge of limited energy availability in embedded devices, this design supports both exact and approximate computations with multiple configurable accuracy levels via a dedicated control register. This feature enables fine-grained control over the trade-off between energy consumption and computational accuracy within standard processor pipelines. Performance evaluations indicate significant power reductions, ranging from 44% to 52% in exact modes and 62% to 68% in approximate modes, while maintaining a computational performance of 1.89 DMIPS/MHz. Further tests on error-tolerant workloads, such as 2D convolution and matrix multiplication, demonstrated up to a 63% reduction in energy consumption. The architecture achieved an efficiency of 1.21 pJ per instruction for matrix multiplication tasks. These results confirm the design's effectiveness for deploying energy-constrained edge AI solutions, offering a robust hardware-level optimization for modern deep learning applications without compromising essential computational capabilities.
Wire timeline
Reconfigurable Multiplier Architecture for Error-Resilient RISC-V Cores
Researchers have introduced a novel runtime reconfigurable multiplier architecture integrated into the RISC-V core, specifically designed to enhance energy efficiency in neural network inference and edge AI applications. Addressing the critical challenge of limited energy availability in embedded devices, this design supports both exact and approximate computations with multiple configurable accuracy levels via a dedicated control register. This feature enables fine-grained control over the trade-off between energy consumption and computational accuracy within standard processor pipelines. Performance evaluations indicate significant power reductions, ranging from 44% to 52% in exact modes and 62% to 68% in approximate modes, while maintaining a computational performance of 1.89 DMIPS/MHz. Further tests on error-tolerant workloads, such as 2D convolution and matrix multiplication, demonstrated up to a 63% reduction in energy consumption. The architecture achieved an efficiency of 1.21 pJ per instruction for matrix multiplication tasks. These results confirm the design's effectiveness for deploying energy-constrained edge AI solutions, offering a robust hardware-level optimization for modern deep learning applications without compromising essential computational capabilities.
cs.AI updates on arXiv.org