With the capacity and performance of FPGAs suitable for space borne applications continuously increasing, the design of FPGAs is becoming increasingly complex involving trading off or simultaneous optimization of space, speed, and power. RNET and ANL are proposing to develop software infrastructure that facilitates automatic performance tuning of FPGAs in terms of speed, power, and size. We introduce an extensible empirical tuning tool system OrFPGA, which is aimed at improving both performance and productivity by enabling FPGA designers to create simple scripts that trigger various FPGA performance optimizations for a specific design. OrFPGA will generate various tuned versions of the same design with different designer parameters and evaluates the versions to select the best performing one for production use. The proposed work will leverage an existing performance tuning tool named Orio developed by ANL for empirical tuning of compute-intensive kernels for a given architecture.
More »