A Survey of Accelerator Architectures for Deep Neural Networks

Modern computing systems are overwhelmingly designed to move data to computation. This design choice goes directly against at least three key trends in computing that cause performance, scalability and energy bottlenecks: (1) data access is a key bottleneck as many important applications are increasingly data-intensive, and memory bandwidth and energy do not scale well, (2) energy consumption is a key limiter in almost all computing platforms, especially server and mobile systems, (3) data movement, especially off-chip to on-chip, is very expensive in terms of bandwidth, energy and latency, much more so than computation. These trends are especially severely-felt in the data-intensive server and energy-constrained mobile systems of today. At the same time, conventional memory technology is facing many technology scaling challenges in terms of reliability, energy, and performance. As a result, memory system architects are open to organizing memory in different ways and making it more intelligent, at the expense of higher cost. The emergence of 3D-stacked memory plus logic, the adoption of error correcting codes inside the latest DRAM chips, proliferation of different main memory standards and chips, specialized for different purposes (e.g., graphics, low-power, high bandwidth, low latency), and the necessity of designing new solutions to serious reliability and security issues, such as the RowHammer phenomenon, are an evidence of this trend. This chapter discusses recent research that aims to practically enable computation close to data, an approach we call processing-in-memory (PIM). PIM places computation mechanisms in or near where the data is stored (i.e., inside the memory chips, in the logic layer of 3D-stacked memory, or in the memory controllers), so that data movement between the computation units and memory is reduced or eliminated.

A Modern Primer on Processing in Memory.

Racetrack memory (RTM) is a novel spintronic memory-storage technology that has the potential to overcome fundamental constraints of existing memory and storage devices. It is unique in that its core differentiating feature is the movement of data, which is composed of magnetic domain walls (DWs), by short current pulses. This enables more data to be stored per unit area compared to any other current technologies. On the one hand, RTM has the potential for mass data storage with unlimited endurance using considerably less energy than today’s technologies. On the other hand, RTM promises an ultrafast nonvolatile memory competitive with static random access memory (SRAM) but with a much smaller footprint. During the last decade, the discovery of novel physical mechanisms to operate RTM has led to a major enhancement in the efficiency with which nanoscopic, chiral DWs can be manipulated. New materials and artificially atomically engineered thin-film structures have been found to increase the speed and lower the threshold current with which the data bits can be manipulated. With these recent developments, RTM has attracted the attention of the computer architecture community that has evaluated the use of RTM at various levels in the memory stack. Recent studies advocate RTM as a promising compromise between, on the one hand, power-hungry, volatile memories and, on the other hand, slow, nonvolatile storage. By optimizing the memory subsystem, significant performance improvements can be achieved, enabling a new era of cache, graphical processing units, and high capacity memory devices. In this article, we provide an overview of the major developments of RTM technology from both the physics and computer architecture perspectives over the past decade. We identify the remaining challenges and give an outlook on its future.

/pdf/magnetic-racetrack-memory-from-physics-to-the-cusp-of-5brh05qnrb.pdf

Magnetic Racetrack Memory: From Physics to the Cusp of Applications Within a Decade

Large deep neural network (DNN) models pose the key challenge to energy efficiency due to the significantly higher energy consumption of off-chip DRAM accesses than arithmetic or SRAM operations. It motivates the intensive research on model compression with two main approaches. Weight pruning leverages the redundancy in the number of weights and can be performed in a non-structured, which has higher flexibility and pruning rate but incurs index accesses due to irregular weights, or structured manner, which preserves the full matrix structure with lower pruning rate. Weight quantization leverages the redundancy in the number of bits in weights. Compared to pruning, quantization is much more hardware-friendly, and has become a "must-do" step for FPGA and ASIC implementations. This paper provides a definitive answer to the question for the first time. First, we build ADMM-NN-S by extending and enhancing ADMM-NN, a recently proposed joint weight pruning and quantization framework. Second, we develop a methodology for fair and fundamental comparison of non-structured and structured pruning in terms of both storage and computation efficiency. Our results show that ADMM-NN-S consistently outperforms the prior art: (i) it achieves 348x, 36x, and 8x overall weight pruning on LeNet-5, AlexNet, and ResNet-50, respectively, with (almost) zero accuracy loss; (ii) we demonstrate the first fully binarized (for all layers) DNNs can be lossless in accuracy in many cases. These results provide a strong baseline and credibility of our study. Based on the proposed comparison framework, with the same accuracy and quantization, the results show that non-structrued pruning is not competitive in terms of both storage and computation efficiency. Thus, we conclude that non-structured pruning is considered harmful. We urge the community not to continue the DNN inference acceleration for non-structured sparsity.

Non-Structured DNN Weight Pruning -- Is It Beneficial in Any Platform?

In‐memory computing is a computing scheme that integrates data storage and arithmetic computation functions. Resistive random access memory (RRAM) arrays with innovative peripheral circuitry provide the capability of performing vector‐matrix multiplication beyond the basic Boolean logic. With such a memory–computation duality, RRAM‐based in‐memory computing enables an efficient hardware solution for matrix‐multiplication‐dependent neural networks and related applications. Herein, the recent development of RRAM nanoscale devices and the parallel progress on circuit and microarchitecture layers are discussed. Well suited for analog synapse and neuron implementation, RRAM device properties and characteristics are emphasized herein. 3D‐stackable RRAM and on‐chip training are introduced in large‐scale integration. The circuit design and system organization of RRAM‐based in‐memory computing are essential to breaking the von Neumann bottleneck. These outcomes illuminate the way for the large‐scale implementation of ultra‐low‐power and dense neural network accelerators.

https://onlinelibrary.wiley.com/doi/epdf/10.1002/aisy.201900068

Resistive Memory‐Based In‐Memory Computing: From Device and Large‐Scale Integration System Perspectives

As a powerful unsupervised learning method, Generative Adversarial Network (GAN) plays an important role in many domains such as video prediction and autonomous driving. It is one of the ten breakthrough technologies in 2018 reported in MIT Technology Review. However, training a GAN imposes three more challenges: (1) intensive communication caused by complex train phases of GAN, (2) much more ineffectual computations caused by special convolutions, and (3) more frequent off-chip memory accesses for exchanging inter-mediate data between the generator and the discriminator. In this paper, we propose LerGAN1, a PIM-based GAN accelerator to address the challenges of training GAN. We first propose a zero-free data reshaping scheme for ReRAM-based PIM, which removes the zero-related computations. We then propose a 3D-connected PIM, which can reconfigure connections inside PIM dynamically according to dataflows of propagation and updating. Our proposed techniques reduce data movement to a great extent, avoiding I/O to become a bottleneck of training GANs. Finally, we propose LerGAN based on these two techniques, providing different levels of accelerating GAN for programmers. Experiments shows that LerGAN achieves 47.2x, 21.42x and 7.46x speedup over FPGA-based GAN accelerator, GPU platform, and ReRAM-based neural network accelerator respectively. Moreover, LerGAN achieves 9.75x, 7.68x energy saving on average over GPU platform, ReRAM-based neural network accelerator respectively, and has 1.04x energy consuming over FPGA-based GAN accelerator.

LerGAN: a zero-free, low data movement and PIM-based GAN architecture

Scratchpad Memory (SPM) has been widely adopted in various computing systems to improve performance of data access. Recently, non-volatile memory technologies (NVMs) have been employed for SPM design to improve its capacity and reduce its energy consumption. In this paper, we explore data allocation in SPM based on racetrack memory (RM), which is an emerging NVM with ultra-high storage density and fast access speed. Since a shift operation is needed to access data in RM, data allocation has an impact on performance of RM based SPM. Several allocation methods have been discussed and compared in this work. Especially, we addressed how to leverage genetic algorithm to achieve near-optimal data allocation.

/pdf/exploring-data-placement-in-racetrack-memory-based-ie6im8tif0.pdf

Exploring data placement in racetrack memory based scratchpad memory

Read mapping is a fundamental step in many genomics applications. It is used to identify potential matches and differences between fragments (called reads) of a sequenced genome and an already known genome (called a reference genome). Read mapping is costly because it needs to perform approximate string matching (ASM) on large amounts of data. To address the computational challenges in genome analysis, many prior works propose various approaches such as accurate filters that select the reads within a dataset of genomic reads (called a read set) that must undergo expensive computation, efficient heuristics, and hardware acceleration. While effective at reducing the amount of expensive computation, all such approaches still require the costly movement of a large amount of data from storage to the rest of the system, which can significantly lower the end-to-end performance of read mapping in conventional and emerging genomics systems. We propose GenStore, the first in-storage processing system designed for genome sequence analysis that greatly reduces both data movement and computational overheads of genome sequence analysis by exploiting low-cost and accurate in-storage filters. GenStore leverages hardware/software co-design to address the challenges of in-storage processing, supporting reads with 1) different properties such as read lengths and error rates, which highly depend on the sequencing technology, and 2) different degrees of genetic variation compared to the reference genome, which highly depends on the genomes that are being compared. Through rigorous analysis of read mapping processes of reads with different properties and degrees of genetic variation, we meticulously design low-cost hardware accelerators and data/computation flows inside a NAND flash-based solid-state drive (SSD). Our evaluation using a wide range of real genomic datasets shows that GenStore, when implemented in three modern NAND flash-based SSDs, significantly improves the read mapping performance of state-of-the-art software (hardware) baselines by 2.07-6.05× (1.52-3.32×) for read sets with high similarity to the reference genome and 1.45-33.63× (2.70-19.2×) for read sets with low similarity to the reference genome.

GenStore: a high-performance in-storage processing system for genome sequence analysis

Data encryption and authentication are essential for secure NVM. However, the introduced security metadata needs to be atomically written back to NVM along with data, so as to provide crash consistency, which unfortunately incurs high overhead. To support fine-grained data protection without compromising the performance, we propose cc-NVM. It firstly proposes an epoch-based mechanism to aggressively cache the security metadata in CPU cache while retaining the consistency of them in NVM. Deferred spreading is also introduced to reduce the calculating overhead for data authentication. Leveraging the hidden ability of data HMACs, we can always recover the consistent but old security metadata to its newest version. Compared to Osiris, a state-of-the-art secure NVM, cc-NVM improves performance by 20.4% on average. When the system crashes, instead of dropping all the data due to malicious attacks, cc-NVM is able to detect and locate the exact tampered data while only incurring extra write traffic by 29.6% on average.

No Compromises: Secure NVM with Crash Consistency, Write-Efficiency and High-Performance

Read mapping is a fundamental, yet computationally-expensive step in many genomics applications. It is used to identify potential matches and differences between fragments (called reads) of a sequenced genome and an already known genome (called a reference genome). To address the computational challenges in genome analysis, many prior works propose various approaches such as filters that select the reads that must undergo expensive computation, efficient heuristics, and hardware acceleration. While effective at reducing the computation overhead, all such approaches still require the costly movement of a large amount of data from storage to the rest of the system, which can significantly lower the end-to-end performance of read mapping in conventional and emerging genomics systems. We propose GenStore, the first in-storage processing system designed for genome sequence analysis that greatly reduces both data movement and computational overheads of genome sequence analysis by exploiting low-cost and accurate in-storage filters. GenStore leverages hardware/software co-design to address the challenges of in-storage processing, supporting reads with 1) different read lengths and error rates, and 2) different degrees of genetic variation. Through rigorous analysis of read mapping processes, we meticulously design low-cost hardware accelerators and data/computation flows inside a NAND flash-based SSD. Our evaluation using a wide range of real genomic datasets shows that GenStore, when implemented in three modern SSDs, significantly improves the read mapping performance of state-of-the-art software (hardware) baselines by 2.07-6.05$\times$ (1.52-3.32$\times$) for read sets with high similarity to the reference genome and 1.45-33.63$\times$ (2.70-19.2$\times$) for read sets with low similarity to the reference genome.

Haiyu Mao

Papers

LerGAN: a zero-free, low data movement and PIM-based GAN architecture

Exploring data placement in racetrack memory based scratchpad memory

GenStore: a high-performance in-storage processing system for genome sequence analysis

No Compromises: Secure NVM with Crash Consistency, Write-Efficiency and High-Performance

GenStore: A High-Performance and Energy-Efficient In-Storage Computing System for Genome Sequence Analysis