The Potential for a GPU-Like Overlay Architecture for FPGAs

We propose a soft processor programmingmodel and architecture inspired by graphics processing units(GPUs) that are well-matched to the strengths of FPGAs,namely, highly parallel and pipelinable computation. Inparticular, our soft processor architecture exploits multithreading,vector operations, and predication to supply afloating-point pipeline of 64 stages via hardware supportfor up to 256 concurrent thread contexts. The key newcontributions of our architecture are mechanisms for managingthreads and register files that maximize data-level andinstruction-level parallelism while overcoming the challengesof port limitations of FPGA block memories as well asmemory and pipeline latency. Through simulation of asystem that (i) is programmable via NVIDIA's high-levelCg language, (ii) supports AMD's CTM r5xx GPU ISA, and(iii) is realizable on an XtremeData XD1000 FPGA-basedaccelerator system, we demonstrate the potential for sucha system to achieve 100% utilization of a deeply pipelinedfloating-point datapath