Abstract: Radio based communication systems and imagers operate under real-time constraints. Off-loading computes to an FPGA seems like a solution to speeding-up your application but comes with many pitfalls. Specifically, software-oriented implementations fail to achieve the required interface bandwidths or computational throughput required to see a speed-up. In this talk, we will discuss the organization of common compute motif's in software-defined-radio and their complexity in time and resources for FPGAs.
Rough goals of talk: 1) Communicate why FPGA acceleration would be attractive 2) Discuss common pitfalls 2a) A behaviorally oriented accelerator 2b) Starving the beast, failing to provide the required data bandwidths 2c) Processor oriented runtime, creates execution overheads 3) Thinking about accelerators, what do they look like 3a) FFT 3b) Correlators 3c) Matrix-Vector Multiply 4) Building an off-load model
Speakers: John Brunhaver