Loading...
Citations
Altmetric:
Abstract
Embedded and mobile systems must be able to execute a variety of different types of code, often with minimal available hardware. Many embedded systems now come with a simple processor and an FPGA, but not more energy-hungry components, such as a GPGPU. In this dissertation we present FlexGrip, a soft architecture which allows for the execution of GPGPU code on an FPGA without the need to recompile the design. The architecture is optimized for FPGA implementation to effectively support the conditional and thread-based execution characteristics of GPGPU execution without FPGA design recompilation. This architecture supports direct CUDA compilation to a binary which is executable on the FPGA-based GPGPU. Our architecture is customizable, thus providing the FPGA designer with a selection of GPGPU cores which display performance versus area tradeoffs. This dissertation describes the FlexGrip architecture in detail and showcases the benefits by evaluating the design for a collection of five standard CUDA benchmarks which are compiled using standard GPGPU compilation tools. Speedups of 23x, on average, versus a MicroBlaze microprocessor are achieved for designs which take advantage of the conditional execution capabilities offered by FlexGrip. We also show FlexGrip can achieve an 80% average reduction of dynamic energy versus the MicroBlaze microprocessor. The dissertation furthers discussion by exploring application-customized versions of the soft GPGPU, thus exploiting the overlay architecture. We expand the architecture to multiple processors per GPGPU and optimizing away features which are not needed by certain classes of applications. These optimizations, which include the effective use of block RAMs and DSP blocks, are critical to the performance of FlexGrip. By implementing a 2 GPGPU design, we show speedups of 44x on average versus a MicroBlaze microprocessor. Application-customized versions of the soft GPGPU can be used to further reduce dynamic energy consumption by an average of 14%. To complete this thesis, we augmented a GPGPU cycle accurate simulator to emulate FlexGrip and evaluate different levels of cache design spaces. We show performance increases for select benchmarks, however, we also show that 64% and 45% of benchmarks exhibited performance decreases when L1D cache was enabled for the 1 SMP and 2 SMP configurations, and only one benchmark showed performance improvement when the L2 cache was enabled.
Type
dissertation
Date
2018-09