In chiplet-based heterogeneous architectures, electrical network-on-package (NoP) designs are typically over-provisioned with routers and channels to provide sufficient bandwidth during periods of high network load. Observing that there are significant periods of low/idle network utilization, prior work has proposed modified network-on-chip (NoC) architectures to enable in-network compute, especially for compute-intensive operations (e.g. linear algebra). However, electrical package-level interconnects impose fundamental energy and bandwidth scaling issues for future chiplet architectures.

Hardware accelerators provide significant speedup and improve energy efficiency for several demanding deep neural network (DNN) applications. DNNs have several hidden layers that perform concurrent matrix-vector multiplications (MVMs) between the network weights and input features. As MVMs are critical to the performance of DNNs, previous research has optimized the performance and energy efficiency of MVMs at both the architecture and algorithm levels. In this paper, we propose to use emerging silicon photonics technology to improve parallelism, speed and overall efficiency with the goal of providing real-time inference and fast training of neural nets. We use microring resonators (MRRs) and Mach-Zehnder interferometers (MZIs) to design two versions (all-optical and partial-optical) of hybrid matrix multiplications for DNNs. Our results indicate that our partial optical design gave the best performance in both energy efficiency and latency, with a reduction of 33.1% for energy-delay product (EDP) with conservative estimates and a 76.4% reduction for EDP with aggressive estimates.

With the end of Dennard scaling, highly-parallel and specialized hardware accelerators have been proposed to improve the throughput and energy-efficiency of deep neural network (DNN) models for various applications. However, collective data movement primitives such as multicast and broadcast that are required for multiply-and-accumulate (MAC) computation in DNN models are expensive, and require excessive energy and latency when implemented with electrical networks. This consequently limits the scalability and performance of electronic hardware accelerators. Emerging technology such as silicon photonics can inherently provide efficient implementation of multicast and broadcast operations, making photonics more amenable to exploit parallelism within DNN models. Moreover, when coupled with other unique features such as low energy consumption, high channel capacity with wavelength-division multiplexing (WDM), and high speed, silicon photonics could potentially provide a viable technology for scaling DNN acceleration.In this paper, we propose Albireo, an analog photonic architecture for scaling DNN acceleration. By characterizing photonic devices such as microring resonators (MRRs) and Mach-Zehnder modulators (MZM) using photonic simulators, we develop realistic device models and outline their capability for system level acceleration. Using the device models, we develop an efficient broadcast combined with multicast data distribution by leveraging parameter sharing through unique WDM dot product processing. We evaluate the energy and throughput performance of Albireo on DNN models such as ResNet18, MobileNet and VGG16. When compared to cur-rent state-of-the-art electronic accelerators, Albireo increases throughput by 110 X, and improves energy-delay product (EDP) by an average of 74 X with current photonic devices. Furthermore, by considering moderate and aggressive photonic scaling, the proposed Albireo design shows that EDP can be reduced by at least 229 X.

Machine learning (ML) architectures such as Deep Neural Networks (DNNs) have achieved unprecedented accuracy on modern applications such as image classification and speech recognition. With power dissipation becoming a major concern in ML architectures, computer architects have focused on designing both energy-efficient hardware platforms as well as optimizing ML algorithms. To dramatically reduce power consumption and increase parallelism in neural network accelerators, disruptive technology such as silicon photonics has been proposed which can improve the performance-per-Watt when compared to electrical implementation. In this paper, we propose PIXEL - Photonic Neural Network Accelerator that efficiently implements the fundamental operation in neural computation, namely the multiply and accumulate (MAC) functionality using photonic components such as microring resonators (MRRs) and Mach-Zehnder interferometer (MZI). We design two versions of PIXEL - a hybrid version that multiplies optically and accumulates electrically and a fully optical version that multiplies and accumulates optically. We perform a detailed power, area and timing analysis of the different versions of photonic and electronic accelerators for different convolution neural networks (AlexNet, VGG16, and others). Our results indicate a significant improvement in the energy-delay product for both PIXEL designs over traditional electrical designs (48.4% for OE and 73.9% for OO) while minimizing latency, at the cost of increased area over electrical designs.

As the number of processing cores and associated threads in chip multiprocessors (CMPs) continues to scale out, on-chip memory access latency dominates application execution time due to increased data movement. Although tiled CMP architectures with distributed shared caches provide a scalable design, increased physical distance between requesting and responding cores has led to both increased on-chip memory access latency and excess energy consumption. Near data processing is a promising approach that can migrate threads closer to data, however prior hand-engineered rules for fine-grained hardware-level thread migration are either too slow to react to changes in data access patterns, or unable to exploit the large variety of data access patterns. In this article, we propose to use reinforcement learning (RL) to learn relatively complex data access patterns to improve on hardware-level thread migration techniques. By utilizing the recent history of memory access locations as input, each thread learns to recognize the relationship between prior access patterns and future memory access locations. This leads to the unique ability of the proposed technique to make fewer, more effective migrations to intermediate cores that minimize the distance to multiple distinct memory access locations. By allowing a low-overhead RL agent to learn a policy from real interaction with parallel programming benchmarks in a parallel simulator, we show that a migration policy which recognizes more complex data access patterns can be learned. The proposed approach reduces on-chip data movement and energy consumption by an average of 41%, while reducing execution time by 43% when compared to a simple baseline with no thread migration; furthermore, energy consumption and execution time are reduced by an additional 10% when compared to a hand-engineered fine-grained migration policy.