Speaker
Description
In recent years, research in the space community has shown a growing interest in Artificial Intelligence (AI), mostly driven by systems miniaturization and commercial competition. Among the available devices for accelerating AI onboard satellites, Field Programmable Gate Arrays (FPGAs) constitute a valuable solution for their energy efficiency and low non-recurrent costs. To facilitate and accelerate the experimentation and development of AI on FPGAs, several automation tooflows have been released.
This session presents a novel technology-independent toolflow for automating DNN deployment onboard FPGAs in space applications. Given an input DNN, the framework first applies compression techniques to shrink model complexity and ease hardware implementation. The acceleration stage of the proposed system features a fully handcrafted Hardware Description Languages (HDL)-based architecture that poses no limit on device portability thanks to the absence of third-party IPs, high scalability, and fine-grain control on resource mapping. An automation process directly generates the HDL sources of the accelerator customized for the target DNN-FPGA pair, thus making the presented solution an end-to-end and ready-to-use toolflow. The user has a high degree of control over the final design as he can indicate constraints on accuracy, inference time, or resource usage percentages.
The presentation illustrates the design choices behind the system dataflow and also provides an insight into user control. We present and discuss implementation results of DNN models on both radiation tolerant and rad-hardened devices from different vendors (Xilinx, Microsemi).
Thanks to its high device portability, the proposed toolflow is a valuable candidate for the deployment of DNNs onboard devices not yet supported by any other framework. The availability of a DNN-to-FPGA toolflow that fully support state-of-the-art space-qualified FPGAs, including NanoXplore technology, will deeply promote the deployment of AI in space missions.
