Optimizations for energy efficiency in GPGPU architectures

It is commonplace for graphics processing units or GPUs today to render extremely complex 3D scenes and textures, in real time, both in the traditional and mobile computing spaces. The computational power required to do this makes them a valuable resource to exploit for general purpose computation. In order to map programs originally designed for sequential CPUs onto massively parallel GPU architectures, it would be necessary to justify the transition with huge performance benefits. Over the last couple of years, there have been numerous proposals to improve the performance of GPUs used for general purpose computing (GPGPUs), but without much consideration for energy efficiency. In my dissertation, I evaluate the feasibility of GPGPUs from an energy perspective and propose some optimizations based on the unique programming model used by GPGPUs. First, I describe the simulation infrastructure, one of the few available to model GPGPUs today, both individually and as part of a heterogeneous system. Next, I propose a design using a shared translation lookaside buffer (TLB) to eliminate chronic memory copies between the CPU and GPU addressing spaces, making heterogeneous CPU-GPU designs energy efficient. Furthermore, to improve the energy efficiency of the on-chip memory hierarchy, I propose adding tiny incoherent caches per processing element, which can filter out frequent accesses to large shared and energy-inefficient cache structures. Finally, I evaluate a design which moves away from the underlying SIMD architecture of GPUs towards a more MIMD-like architecture, enabling the execution of both CPU and GPGPU workloads without negatively affecting the energy efficiency availed by traditional workloads on GPGPUs