Processing-in-memory (PIM) is a promising solution to address the "memory wall" challenges for future computer systems. Prior proposed PIM architectures put additional computation logic in or near memory. The emerging metal-oxide resistive random access memory (ReRAM) has showed its potential to be used for main memory. Moreover, with its crossbar array structure, ReRAM can perform matrix-vector multiplication efficiently, and has been widely studied to accelerate neural network (NN) applications. In this work, we propose a novel PIM architecture, called PRIME, to accelerate NN applications in ReRAM based main memory. In PRIME, a portion of ReRAM crossbar arrays can be configured as accelerators for NN applications or as normal memory for a larger memory space. We provide microarchitecture and circuit designs to enable the morphable functions with an insignificant area overhead. We also design a software/hardware interface for software developers to implement various NNs on PRIME. Benefiting from both the PIM architecture and the efficiency of using ReRAM for NN computation, PRIME distinguishes itself from prior work on NN acceleration, with significant performance improvement and energy saving. Our experimental results show that, compared with a state-of-the-art neural processing unit design, PRIME improves the performance by ~2360× and the energy consumption by ~895×, across the evaluated machine learning benchmarks.

PRIME: a novel processing-in-memory architecture for neural network computation in ReRAM-based main memory

IEEE transactions on computer-aided design of integrated circuits and systems : a publication of the IEEE Circuits and Systems Society

Deep learning algorithms achieve high classification accuracy at the expense of significant computation cost To address this cost, a number of quantization schemes have been proposed - but most of these techniques focused on quantizing weights, which are relatively smaller in size compared to activations This paper proposes a novel quantization scheme for activations during training - that enables neural networks to work well with ultra low precision weights and activations without any significant accuracy degradation This technique, PArameterized Clipping acTivation (PACT), uses an activation clipping parameter $\alpha$ that is optimized during training to find the right quantization scale PACT allows quantizing activations to arbitrary bit precisions, while achieving much better accuracy relative to published state-of-the-art quantization schemes We show, for the first time, that both weights and activations can be quantized to 4-bits of precision while still achieving accuracy comparable to full precision networks across a range of popular models and datasets We also show that exploiting these reduced-precision computational units in hardware can enable a super-linear improvement in inferencing performance due to a significant reduction in the area of accelerator compute engines coupled with the ability to retain the quantized model and activation data in on-chip memories

/pdf/pact-parameterized-clipping-activation-for-quantized-neural-544x8orkhr.pdf

PACT: Parameterized Clipping Activation for Quantized Neural Networks

Convolution neural networks (CNNs) are the heart of deep learning applications. Recent works PRIME [1] and ISAAC [2] demonstrated the promise of using resistive random access memory (ReRAM) to perform neural computations in memory. We found that training cannot be efficiently supported with the current schemes. First, they do not consider weight update and complex data dependency in training procedure. Second, ISAAC attempts to increase system throughput with a very deep pipeline. It is only beneficial when a large number of consecutive images can be fed into the architecture. In training, the notion of batch (e.g. 64) limits the number of images can be processed consecutively, because the images in the next batch need to be processed based on the updated weights. Third, the deep pipeline in ISAAC is vulnerable to pipeline bubbles and execution stall. In this paper, we present PipeLayer, a ReRAM-based PIM accelerator for CNNs that support both training and testing. We analyze data dependency and weight update in training algorithms and propose efficient pipeline to exploit inter-layer parallelism. To exploit intra-layer parallelism, we propose highly parallel design based on the notion of parallelism granularity and weight replication. With these design choices, PipeLayer enables the highly pipelined execution of both training and testing, without introducing the potential stalls in previous work. The experiment results show that, PipeLayer achieves the speedups of 42.45x compared with GPU platform on average. The average energy saving of PipeLayer compared with GPU implementation is 7.17x.

PipeLayer: A Pipelined ReRAM-Based Accelerator for Deep Learning

High-performance parallel computer architecture and systems have been improved at a phenomenal rate. In the meantime, VLSI computer-aided design (CAD) software for multibillion-transistor IC design has become increasingly complex and requires prohibitively high computational resources. Recent studies have shown that, numerous CAD problems, with their high computational complexity, can greatly benefit from the fast-increasing parallel computation capabilities. However, parallel programming imposes big challenges for CAD applications. Fully exploiting the computational power of emerging general-purpose and domain-specific multicore/many-core processor systems, calls for fundamental research and engineering practice across every stage of parallel CAD design, from algorithm exploration, programming models, design-time and run-time environment, to CAD applications, such as verification, optimization, and simulation. This journal special section will cover recent progress on parallel CAD research, including algorithm foundations, programming models, parallel architectural-specific optimization, and verification. More specifically, papers with in-depth and extensive coverage of the following topics will be considered, as well as other topics relevant to the design of parallel CAD algorithms and software tools. 1. Parallel algorithm design and specification for CAD applications 2. Parallel programming models and languages of particular use in CAD 3. Runtime support and performance optimization for CAD applications 4. Parallel architecture-specific design and optimization for CAD applications 5. Parallel program debugging and verification techniques particularly relevant for CAD The papers should be submitted via the Manuscript Central website and should adhere to standard ACM TODAES formatting requirements (http://todaes.acm.org/). The page count limit is 25.

Parallel CAD: Algorithm Design and Programming Special Section Call for Papers TODAES: ACM Transactions on Design Automation of Electronic Systems

The explosion of digital data and the ever-growing need for fast data analysis have made in-memory big-data processing in computer systems increasingly important. In particular, large-scale graph processing is gaining attention due to its broad applicability from social science to machine learning. However, scalable hardware design that can efficiently process large graphs in main memory is still an open problem. Ideally, cost-effective and scalable graph processing systems can be realized by building a system whose performance increases proportionally with the sizes of graphs that can be stored in the system, which is extremely challenging in conventional systems due to severe memory bandwidth limitations. In this work, we argue that the conventional concept of processing-in-memory (PIM) can be a viable solution to achieve such an objective. The key modern enabler for PIM is the recent advancement of the 3D integration technology that facilitates stacking logic and memory dies in a single package, which was not available when the PIM concept was originally examined. In order to take advantage of such a new technology to enable memory-capacity-proportional performance, we design a programmable PIM accelerator for large-scale graph processing called Tesseract. Tesseract is composed of (1) a new hardware architecture that fully utilizes the available memory bandwidth, (2) an efficient method of communication between different memory partitions, and (3) a programming interface that reflects and exploits the unique hardware design. It also includes two hardware prefetchers specialized for memory access patterns of graph processing, which operate based on the hints provided by our programming model. Our comprehensive evaluations using five state-of-the-art graph processing workloads with large real-world graphs show that the proposed architecture improves average system performance by a factor of ten and achieves 87% average energy reduction over conventional systems.

/pdf/a-scalable-processing-in-memory-accelerator-for-parallel-23yasy5934.pdf

A scalable processing-in-memory accelerator for parallel graph processing

Processing-in-memory (PIM) is rapidly rising as a viable solution for the memory wall crisis, rebounding from its unsuccessful attempts in 1990s due to practicality concerns, which are alleviated with recent advances in 3D stacking technologies. However, it is still challenging to integrate the PIM architectures with existing systems in a seamless manner due to two common characteristics: unconventional programming models for in-memory computation units and lack of ability to utilize large on-chip caches. In this paper, we propose a new PIM architecture that (1) does not change the existing sequential programming models and (2) automatically decides whether to execute PIM operations in memory or processors depending on the locality of data. The key idea is to implement simple in-memory computation using compute-capable memory commands and use specialized instructions, which we call PIM-enabled instructions, to invoke in-memory computation. This allows PIM operations to be interoperable with existing programming models, cache coherence protocols, and virtual memory mechanisms with no modification. In addition, we introduce a simple hardware structure that monitors the locality of data accessed by a PIM-enabled instruction at runtime to adaptively execute the instruction at the host processor (instead of in memory) when the instruction can benefit from large on-chip caches. Consequently, our architecture provides the illusion that PIM operations are executed as if they were host processor instructions. We provide a case study of how ten emerging data-intensive workloads can benefit from our new PIM abstraction and its hardware implementation. Evaluations show that our architecture significantly improves system performance and, more importantly, combines the best parts of conventional and PIM architectures by adapting to data locality of applications.

/pdf/pim-enabled-instructions-a-low-overhead-locality-aware-4gcijahtny.pdf

PIM-enabled instructions: a low-overhead, locality-aware processing-in-memory architecture

Quantization is considered as one of the most effective methods to optimize the inference cost of neural network models for their deployment to mobile and embedded systems, which have tight resource constraints. In such approaches, it is critical to provide low-cost quantization under a tight accuracy loss constraint (e.g., 1%). In this paper, we propose a novel method for quantizing weights and activations based on the concept of weighted entropy. Unlike recent work on binary-weight neural networks, our approach is multi-bit quantization, in which weights and activations can be quantized by any number of bits depending on the target accuracy. This facilitates much more flexible exploitation of accuracy-performance trade-off provided by different levels of quantization. Moreover, our scheme provides an automated quantization flow based on conventional training algorithms, which greatly reduces the design-time effort to quantize the network. According to our extensive evaluations based on practical neural network models for image classification (AlexNet, GoogLeNet and ResNet-50/101), object detection (R-FCN with 50-layer ResNet), and language modeling (an LSTM network), our method achieves significant reductions in both the model size and the amount of computation with minimal accuracy loss. Also, compared to existing quantization schemes, ours provides higher accuracy with a similar resource constraint and requires much lower design effort.

https://openaccess.thecvf.com/content_cvpr_2017/papers/Park_Weighted-Entropy-Based_Quantization_for_CVPR_2017_paper.pdf

Weighted-Entropy-Based Quantization for Deep Neural Networks

Spin-Transfer Torque RAM (STT-RAM) has been considered as a promising candidate for on-chip last-level caches, replacing SRAM for better energy efficiency, smaller die footprint, and scalability. However, it also introduces several new challenges into last-level cache design that need to be overcome for feasible deployment of STT-RAM caches. Among other things, mitigating the impact of slow and energy-hungry write operations is of the utmost importance. In this paper, we propose a new mechanism to reduce write activities of STT-RAM last-level caches. The key observation is that a significant amount of data written to last-level caches is not actually re-referenced again during the lifetime of the corresponding cache blocks. Such write operations, which we call dead writes, can bypass the cache without incurring extra misses by definition. Based on this, we propose Dead Write Prediction Assisted STT-RAM Cache Architecture (DASCA), which predicts and bypasses dead writes for write energy reduction. For this purpose, we first propose a novel classification of dead writes, which is composed of dead-on-arrival fills, dead-value fills, and closing writes, as a theoretical model for redundant write elimination. On top of that, we present a dead write predictor based on a state-of-the-art dead block predictor. Evaluations show that our architecture achieves an energy reduction of 68% (62%) in last-level caches and an additional energy reduction of 10% (16%) in main memory and even improves system performance by 6% (14%) on average compared to the STT-RAM baseline in a single-core (quad-core) system.

DASCA: Dead Write Prediction Assisted STT-RAM Cache Architecture

Editor’s note: It has been observed that the majority of the kernel weights and input activations in the state-of-the-art convolution neural networks (CNNs) have zero values. This article proposes a CNN hardware accelerator that exploits this property to achieve significant performance and energy improvements. —Mustafa Ozdal, Bilkent University

Junwhan Ahn

Papers

A scalable processing-in-memory accelerator for parallel graph processing

PIM-enabled instructions: a low-overhead, locality-aware processing-in-memory architecture

Weighted-Entropy-Based Quantization for Deep Neural Networks

DASCA: Dead Write Prediction Assisted STT-RAM Cache Architecture

ZeNA: Zero-Aware Neural Network Accelerator