FPGA based in-memory AI computing

Abstract

The advent of AI in vehicles of all kinds is simultaneously creating the need for more and most often also very large computing capacities. Depending on the type of vehicle, this gives rise to various problems: while overall hardware and engineering costs dominate for airplanes, in fully electrical cars the costs for computing hardware are more of a matter. Common in both domains are tight requirements on the size, weight and space of the hardware, especially for drones and satellites, where this is most challenging. For airplanes and especially for satellites, an additional challenge is the radiation resistance of the usually very memory-intensive AI systems. We therefore propose an FPGA-based in-memory AI computation methodology, which is so far only applicable for small AI systems, but works exclusively with the local memory elements of FPGAs: lookup tables (LUTs) and registers. By not using external and thus slow, inefficient and radiation-sensitive DRAM, but only local SRAM, we can make AI systems faster, lighter and more efficient than is possible with conventional GPUs or AI accelerators. All known radiation hardening techniques for FPGAs also work for our systems.