Performance landscape of resource-constrained platforms targeting DNNs

Abstract: Over the recent years, a significant number of complex, deep neural networks have been developed for a variety of applications including speech and face recognition, computer vision in the areas of health-care, automatic translation, image classification, etc. Moreover, there is an increasing demand in deploying these networks in resource-constrained edge devices. As the computational demands of these models keep increasing, pushing to their limits the targeted devices, the constant development of new hardware systems tailored to those workloads has been observed. Since programmability of these diverse and complex platforms -- compounded by the rapid development of new DNN models -- is a major challenge, platform vendors have developed Machine Learning tailored SDKs to maximize the platform's performance. This work investigates the performance achieved on a number of modern commodity embedded platforms coupled with the vendors' provided software support when state-of-the-art DNN models from image classification, object detection and image segmentation are targeted. The work quantifies the relative latency gains of the particular embedded platforms and provides insights on the relationship between the required minimum batch size for achieving maximum throughput, concluding that modern embedded systems reach their maximum performance even for modest batch sizes when a modern state of the art DNN model is targeted. Overall, the presented results provide a guide for the expected performance for a number of state-of-the-art DNNs on popular embedded platforms across the image classification, detection and segmentation domains.