Standard convolutional neural networks (CNNs) have large amounts of data redundancy, and the same accuracy can be obtained even in lower bit weights instead of floating-point representation. Most CNNs have to be developed and executed on high-end GPU-based workstations, for which it is hard to transplant the existing implementations onto portable edge FPGAs because of the limitation of on-chip block memory storage size and battery capacity. In this paper, we present adaptive pointwise convolution and 2D convolution joint network (AP2D-Net), an ultra-low power and relatively high throughput system combined with dynamic precision weights and activation. Our system has high performance, and we make a trade-off between accuracy and power efficiency by adopting unmanned aerial vehicle (UAV) object detection scenarios. We evaluate our system on the Zynq UltraScale+ MPSoC Ultra96 mobile FPGA platform. The target board can get the real-time speed of 30 fps under 5.6 W, and the FPGA on-chip power is only 0.6 W. The power efficiency of our system is 2.8× better than the best system design on a Jetson TX2 GPU and 1.9× better than the design on a PYNQ-Z1 SoC FPGA.

Clock gating is an effective way to reduce the dynamic power in digital sequential circuits. In this paper, a gate-level activity correlation-based clustering clock-gating (CCG) technique is proposed for digital filters. The CCG technique exploits the correlations between flip-flops, and determines how to group the flip-flops for clock gating. An Activity Correlation Matrix (ACMtx) is introduced to describe the correlations between the flip-flops, and a greedy clustering algorithm is proposed to find an optimised clustering scheme as well. Experiments on ISCAS’89 benchmarks show that the proposed technique can reduce power consumption by 5.08% on average, on top of existing technique. For the circuits with large numbers of flip-flops, our proposed technique can save 15.84% more power on average.

Since hash functions are cryptography's most widely used primitives, efficient hardware implementation of hash functions is of critical importance. The proposed high performance hardware implementation of the hash functions used sponge construction which generates desired length digest, considering two key design metrics: throughput and power consumption. Firstly, this paper introduces unfolding transformation which increases the throughput of hash function and pipelining and parallelism design techniques which reduce the delay. Secondly, we propose a frequency trade-off technique which can give us a scope of frequency value for making a trade-off between low dynamic power consumption and high throughput. Finally, we use load-enable based clock gating scheme to eliminate wasted toggle rate of signals in the idle mode of hash encryption system. We demonstrated the proposed design techniques by using 45 nm CMOS technology at 10 MHz. The results show that we can achieve up to 47.97 times higher throughput, 6.31% delay reduction, and 13.65% dynamic power reduction.