An Approximate GEMM Unit for Energy-Efficient Object Detection

Abstract

Edge computing brings artificial intelligence algorithms and graphics processing units closer to data sources, making autonomy and energy-efficient processing vital for their design. Approximate computing has emerged as a popular strategy for energy-efficient circuit design, where the challenge is to achieve the best tradeoff between design efficiency and accuracy. The essential operation in artificial intelligence algorithms is the general matrix multiplication (GEMM) operation comprised of matrix multiplication and accumulation. This paper presents an approximate general matrix multiplication (AGEMM) unit that employs approximate multipliers to perform matrix–matrix operations on four-by-four matrices given in sixteen-bit signed fixed-point format. The synthesis of the proposed AGEMM unit to the 45 nm Nangate Open Cell Library revealed that it consumed only up to 36% of the area and 25% of the energy required by the exact general matrix multiplication unit. The AGEMM unit is ideally suited to convolutional neural networks, which can adapt to the error induced in the computation. We evaluated the AGEMM units’ usability for honeybee detection with the YOLOv4-tiny convolutional neural network. The results implied that we can deploy the AGEMM units in convolutional neural networks without noticeable performance degradation. Moreover, the AGEMM unit’s employment can lead to more area- and energy-efficient convolutional neural network processing, which in turn could prolong sensors’ and edge nodes’ autonomy.

A GPU-Parallel Image Coregistration Algorithm for InSar Processing at the Edge

“…As future work, we plan to improve the algorithm to implement it on heterogeneous multicore CPU / GPU architectures, as done in [ 36 ], and to optimize the portable design of memory accesses to avoid unwanted overhead on Jetson boards with low CUDA capabilities. Moreover, thanks to the introduction of Volta architecture on the recent Nvidia Tegra series, the availability of tensor cores opens up to new algorithmic designs [ 37 ]. Those devices deliver half-precision GEMM (General Matrix Multiply) in one clock cycle, consuming low-energy in edge context.…”

In Search of an Accuracy-Tuneable Accelerator Platform for Ubiquitous Computing

“…The core element of the ACTA represents a dedicated Approximate General matrix multiply hardware Unit (AGU) whose accuracy can be changed on the fly. The envisioned AGU is based on the iterative logarithmic product approximation proposed by Babić et al [13] and the design of an approximate GEMM unit presented by Pilipović et al [14]. The AGU does not duplicate the functionality in multiple accuracy versions but incorporates a simple logic to approximate addition and multiplication constituting the GEMM operation.…”

Energy-efficient neural network learning with accuracy-adjustable floating-point multiplier

“…There have been several attempts to use approximate integer multipliers in neural network learning [12]- [14]. The authors of these studies report that the learning was successful, but they mainly worked with tiny neural networks.…”