On-chip photonics is a disruptive technology, and such NoCs are superior to traditional electrical NoCs in terms of latency, power, and bandwidth. Hence, researchers have proposed a wide variety of optical networks for multicore processors. The high bandwidth and low latency features of photonic NoCs have led to the overall improvement in the system performance. However, there are very few proposals that discuss the usage of optical interconnects in Graphics Processor Units (GPUs). GPUs can also substantially gain from such novel technologies, because they need to provide significant computational throughput without further stressing their power budgets. The main shortcoming of optical networks is their high static power usage, because the lasers are turned on all the time by default, even when there is no traffic inside the chip, and thus sophisticated laser modulation schemes are required. Such modulation schemes base their decisions on an accurate prediction of network traffic in the future. In this article, we propose an energy-efficient and scalable optical interconnect for modern GPUs called GPUOPT that smartly creates an overlay network by dividing the symmetric multiprocessors (SMs) into clusters. It furthermore has separate sub-networks for coherence and non-coherence traffic. To further increase the throughput, we connect the off-chip memory with optical links as well. Subsequently, we show that traditional laser modulation schemes (for reducing static power consumption) that were designed for multicore processors are not that effective for GPUs. Hence, there was a need to create a bespoke scheme for predicting the laser power usage in GPUs. Using this set of techniques, we were able to improve the performance of a modern GPU by 45% as compared to a state-of-the-art electrical NoC. Moreover, as compared to competing optical NoCs for GPUs, our scheme reduces the laser power consumption by 67%, resulting in a net 65% reduction in ED 2 for a suite of Rodinia benchmarks.

中文翻译：

---

GPUOPT

片上光子学是一种颠覆性技术，此类 NoC 在延迟、功率和带宽方面优于传统的电子 NoC。因此，研究人员提出了多种用于多核处理器的光网络。光子 NoC 的高带宽和低延迟特性导致了系统性能的整体提升。然而，很少有提案讨论在图形处理器单元 (GPU) 中使用光学互连。GPU 也可以从这些新技术中获得实质性收益，因为它们需要提供显着的计算吞吐量，而不会进一步强调其功率预算。光网络的主要缺点是它们的静态功耗很高，因为默认情况下激光器一直打开，即使芯片内部没有流量，因此需要复杂的激光调制方案。这种调制方案的决策基于对未来网络流量的准确预测。在本文中，我们提出了一种用于现代 GPU 的节能且可扩展的光学互连，称为GPUOPT它通过将对称多处理器 (SM) 划分为集群来巧妙地创建覆盖网络。此外，它还具有用于相干和非相干业务的单独子网络。为了进一步提高吞吐量，我们还将片外存储器与光链路连接起来。随后，我们展示了为多核处理器设计的传统激光调制方案（用于降低静态功耗）对 GPU 并不那么有效。因此，需要创建一个定制方案来预测 GPU 中的激光功率使用情况。与最先进的电气 NoC 相比，使用这套技术，我们能够将现代 GPU 的性能提高 45%。此外，与用于 GPU 的竞争光学 NoC 相比，我们的方案将激光功耗降低了 67%，从而净减少了 65%ED 2一套罗迪尼亚基准。