Inference and Evaluation of Deep Convolutional Neural Networks on Microchip's Hardware Accelerator VectorBlox

The exponential growth of artificial intelligence in recent years has created new opportunities across various industries, including the space sector. The increasing significance of CNNs has played a pivotal role in shaping modern AI advancements. As CNNs become more intricate, there arises a pressing need for efficient and automated toolflows to deploy them. FPGA-based solutions offer a promising avenue for acceleration due to their balanced performance, power efficiency, and programmability. This relevance is even more pronounced when considering space applications, where radiation-tolerant FPGAs can play a pivotal role in supporting AI tasks within the unique challenges of the space environment. Toolflows provide automation for intricate design tasks, substantially reducing complexity and effort. Within this context, the focus of this paper is to conduct an exploration of CNN-to-FPGA toolflows, with a particular emphasis on VectorBlox, a toolflow developed by Microchip. The study aims to conduct a comparative analysis between VectorBlox and similar toolflows, evaluating essential metrics to assess its performance, adaptability, and efficiency. Afterward, some tests were conducted to evaluate VectorBlox's performance. First, the inference times of some of the most common convolutional and fully connected neural network patterns are evaluated. Next, a comprehensive analysis of the inference of four NNs is reported, three of which were created by transfer learning with some of the best-known deep CNNs architectures (MobileNet, ResNet, and Inception), and trained on the EuroSAT dataset from ESA's Sentinel 2 mission. The concluding section presents a summary of the major limitations encountered when trying to infer unsupported nns.