The Potential for a GPU-Like Overlay Architecture for FPGAs

We propose a soft processor programming model and architecture inspired by graphics processing units (GPUs) that are well-matched to the strengths of FPGAs, namely, highly parallel and pipelinable computation. In particular, our soft processor architecture exploits multithreading, vector operations, and predication to supply a floating-point pipeline of 64 stages via hardware support for up to 256 concurrent thread contexts. The key new contributions of our architecture are mechanisms for managing threads and register files that maximize data-level and instruction-level parallelism while overcoming the challenges of port limitations of FPGA block memories as well as memory and pipeline latency. Through simulation of a system that (i) is programmable via NVIDIA's high-level Cg language, (ii) supports AMD's CTM r5xx GPU ISA, and (iii) is realizable on an XtremeData XD1000 FPGA-based accelerator system, we demonstrate the potential for such a system to achieve 100% utilization of a deeply pipelined floating-point datapath. 1. Introduction As FPGAs become increasingly dense and powerful, with high-speed I/Os, hard multipliers, and plentiful memory blocks, they have consequently become more desirable platforms for computing. Recently there is building interest in using FPGAs as accelerators for high-performance computing, leading to commercial products such as the SGI RASC which integrates FPGAs into a blade server platform, and XtremeData and Nallatech that offer FPGA accelerator modules that can be installed alongside a conventional CPU in a standard dual-socket motherboard. The challenge for such systems is to provide a programming model that is easily accessible for the programmers in the scientific, financial, and other data-driven arenas that will use them. Developing an accelerator design in a hardware description language such as Verilog is difficult, requiring an expert hardware designer to perform all of the implementation, testing, and debugging required for developing real hardware. Behavioral synthesis techniques—that allow a programmer to write code in a high-level language such as C that is then automatically translated into custom hardware circuits—have long-term promise [1–3], but currently have many limitations. What is needed is a high-level programming model specifically tailored to making the creation of custom FPGA-based accelerators easy. In contrast with the approaches of custom hardware and behavioral synthesis, a more familiar model is to use a standard high-level language and environment to program a processor, or in this case an