Spiking neural networks (SNNs) circumvent the need for large scale arithmetic using techniques inspired by biology.However, SNNs are designed with fundamentally different algorithms from ANNs, which have benefited from a rich history of theoretical advances and an increasingly mature software stack.In this paper we explore the potential of a new technique that lies between these two approaches, one that can leverage the software and system level optimizations of ANNs while utilizing biologically inspired circuits for energy efficient computation.The resulting hardware represents the traditional weight of an ANN as nothing more than a delay element and the degree of activation as nothing more than arrival time of a digital signal.Building on these fundamental operations, we can implement complete ANNs through several innovations: spatial and temporal reuse that facilitates classical dataflows, reducing memory system demands for ANN temporal operations; a new noise-tolerant temporal summation operation; novel hybrid digital/temporal memories; and the integration of temporal memory circuits for shepherding inter-layer activations.Using the MLPerf Tiny benchmark suite, we demonstrate how several architectural parameters can impact inference accuracy, that our proposed systolic array can provide 11× better energy consumption with a 4× improvement in latency compared to SNNs, and when equipped with temporal memories provides 3.5× improvements in energy compared to the most aggressive 8-bit digital systolic arrays.

Convolution is an important operation at the heart of many applications, including image processing, object detection, and neural networks. While data movement and coordination operations continue to be important areas for optimization in general-purpose architectures, for computation fused with sensor operation, the underlying multiply-accumulate (MAC) operations dominate power consumption. Non-traditional data encoding has been shown to reduce the energy consumption of this arithmetic, with options including everything from reduced-precision floating point to fully stochastic operation, but all of these approaches start with the assumption that a complete analog-to-digital conversion (ADC) has already been done for each pixel. While analog-to-time converters have been shown to use less energy, arithmetically manipulating temporally encoded signals beyond simple min, max, and delay operations has not previously been possible, meaning operations such as convolution have been out of reach. In this paper we show that arithmetic manipulation of temporally encoded signals is possible, practical to implement, and extremely energy efficient.