Artificial Intelligence

The increased number of IoT end devices places new demands on computational architectures. The devices (e.g. Wearables, Virtual Reality (VR), Smartphones, Smartcams, Tablelts, PC, etc) stream a large amount of input data and are ideal targets for machine learning applications.

Deep Convolutional Neural Networks (DCNNs) have shown tremendous results for various tasks in computer vision. Methods of improving the performance accuracy and execution time of popular Convolutional Neural Networks (CNNs) have been advanced in several works through load distribution across multiple GPUs. More recent demands for deep learning however have explored power efficient load distribution among computing architectures and this has drawn more attention to FPGAs.

UltraScale FPGAs operate on as low as 25W and have competed closely with high-performing GPU platforms consuming 235W of power. Modern FPGAs designs have also started exploring smaller feature sizes for much dense circuits and also incorporate hardened computational units "SoC FPGAs" together with the generic FPGA fabric which has made it much more possible for deep neural networks to be implemented onto a single FPGA system.

New advances have exploited multiple computation units on single FPGA architectures and different parallelisms in execution which has also significantly improved their performance. NS-T® developers introduced a software defined cloud-edge-end architecture to improve the performance of FPGA based CNNs. The distributed computing hierarchy between the cloud, edge (fog) and end devices minimizes communication and resource utilization on all devices and takes maximum advantage of extracted features to be utilized in the cloud. This also ensures system automated sensor fusion and fault tolerance.

#NS-T for AI