Historically, improvements in GPU-based high performance computing have been tightly coupled to transistor scaling. As Moore's law slows down, and the number of transistors per die no longer grows at historical rates, the performance curve of single monolithic GPUs will ultimately plateau. However, the need for higher performing GPUs continues to exist in many domains. To address this need, in this paper we demonstrate that package-level integration of multiple GPU modules to build larger logical GPUs can enable continuous performance scaling beyond Moore's law. Specifically, we propose partitioning GPUs into easily manufacturable basic GPU Modules (GPMs), and integrating them on package using high bandwidth and power efficient signaling technologies. We lay out the details and evaluate the feasibility of a basic Multi-Chip-Module GPU (MCM-GPU) design. We then propose three architectural optimizations that significantly improve GPM data locality and minimize the sensitivity on inter-GPM bandwidth. Our evaluation shows that the optimized MCM-GPU achieves 22.8% speedup and 5x inter-GPM bandwidth reduction when compared to the basic MCM-GPU architecture. Most importantly, the optimized MCM-GPU design is 45.5% faster than the largest implementable monolithic GPU, and performs within 10% of a hypothetical (and unbuildable) monolithic GPU. Lastly we show that our optimized MCM-GPU is 26.8% faster than an equally equipped Multi-GPU system with the same total number of SMs and DRAM bandwidth.

High performance computing (HPC) applications have parallel code sections that must scale to large numbers of cores, which makes them sensitive to serial regions. Current supercomputing systems with heterogeneous or asymmetric CMPs (ACMP) combine few high-performance big cores for serial regions, together with many low-power lean cores for throughput computing. The low requirements of HPC applications in the core front-end lead some designs, such as SMT and GPU cores, to share front-end structures including the instruction cache (I-cache). However, little work exists to analyze the benefit of sharing the I-cache among full cores, which seems compelling as a solution to reduce silicon area and power. This paper analyzes the performance, power and area impact of such a design on an ACMP with one high-performance core and multiple low-power cores. Having identified that multiple cores run the same code during parallel regions, the lean cores share the I-cache with the intent of benefiting from mutual prefetching, without increasing the average access latency. Our exploration of the multiple parameters finds the sweet spot on a wide interconnect to access the shared I-cache and the inclusion of a few line buffers to provide the required bandwidth and latency to sustain performance. The projections with McPAT and a rich set of HPC benchmarks show 11% area savings with a 5% energy reduction at no performance cost.

GPUs achieve high throughput and power efficiency by employing many small single instruction multiple thread (SIMT) cores. To minimize scheduling logic and performance variance they utilize a uniform memory system and leverage strong data parallelism exposed via the programming model. With Moore's law slowing, for GPUs to continue scaling performance (which largely depends on SIMT core count) they are likely to embrace multi-socket designs where transistors are more readily available. However when moving to such designs, maintaining the illusion of a uniform memory system is increasingly difficult. In this work we investigate multi-socket non-uniform memory access (NUMA) GPU designs and show that significant changes are needed to both the GPU interconnect and cache architectures to achieve performance scalability. We show that application phase effects can be exploited allowing GPU sockets to dynamically optimize their individual interconnect and cache policies, minimizing the impact of NUMA effects. Our NUMA-aware GPU outperforms a single GPU by 1.5×, 2.3×, and 3.2× while achieving 89%, 84%, and 76% of theoretical application scalability in 2, 4, and 8 sockets designs respectively. Implementable today, NUMA-aware multi-socket GPUs may be a promising candidate for scaling GPU performance beyond a single socket.