Low-Area and High-Speed Approximate Matrix-Vector Multiplier

Matrix multiplication is a high-cost operation that can benefit from efficient hardware implementation. Approximate computing offers a promising way to lower the hardware costs and to speed up the computation by leveraging the error tolerance of applications like image processing. This work proposes a novel approximate matrix-vector multiplier which features low area and high speed. Moreover, its accuracy is dynamically reconfigurable, allowing the user to trade small errors for increased speed. Compared to previous designs, our approach reduces the area cost up to 70% with a 5% average error. With a more relaxed 10% error constraint, it achieves a speedup of 2x. We apply the proposed design to color transformation, a basic operation in face-detection algorithms. The approximate transformed images exhibit only a small decrease in detection accuracy.