In response to the escalating demand for hardware-efficient Deep Neural Network (DNN) architectures, we present a novel quantize-enabled Multiply-Accumulate (MAC) unit. Our methodology employs a right shift-and-add computation for MAC operation, enabling runtime truncation without additional hardware. This architecture optimally utilizes hardware resources, enhancing throughput performance while reducing computational complexity through bit-truncation techniques. Our key methodology involves designing a hardware-efficient MAC computational algorithm that supports both iterative and pipeline implementations, catering to diverse hardware efficiency or enhanced throughput requirements in accelerators. Additionally, we introduce a processing element (PE) with a pre-loading bias scheme, reducing one clock delay and eliminating the need for conventional extra resources in PE implementation. The PE facilitates quantization-based MAC calculations through an efficient bit-truncation method, removing the necessity for extra hardware logic. This versatile PE accommodates variable bit-precision with a dynamic fraction part within the sfxpt representation, meeting specific model or layer demands. Through software emulation, our proposed approach demonstrates minimal accuracy loss, revealing under 1.6% loss for LeNet-5 using MNIST and around 4% for ResNet-18 and VGG-16 with CIFAR-10 in the sfxpt<8,5> format compared to conventional float32-based implementations. Hardware performance parameters on the Xilinx-Virtex-7 board unveil a 37% reduction in area utilization and a 45% reduction in power consumption compared to the best state-of-the-art MAC architecture. Extending the proposed MAC to a LeNet DNN model results in a 42% reduction in resource requirements and a significant 27% reduction in delay. This architecture provides notable advantages for resource-efficient, high-throughput edge-AI applications.
