Many architects believe that major improvements in cost-energy-performance must now come from domain-specific hardware. This paper evaluates a custom ASIC---called a Tensor Processing Unit (TPU)---deployed in datacenters since 2015 that accelerates the inference phase of neural networks (NN). The heart of the TPU is a 65,536 8-bit MAC matrix multiply unit that offers a peak throughput of 92 TeraOps/second (TOPS) and a large (28 MiB) software-managed on-chip memory. The TPU's deterministic execution model is a better match to the 99th-percentile response-time requirement of our NN applications than are the time-varying optimizations of CPUs and GPUs (caches, out-of-order execution, multithreading, multiprocessing, prefetching, ...) that help average throughput more than guaranteed latency. The lack of such features helps explain why, despite having myriad MACs and a big memory, the TPU is relatively small and low power. We compare the TPU to a server-class Intel Haswell CPU and an Nvidia K80 GPU, which are contemporaries deployed in the same datacenters. Our workload, written in the high-level TensorFlow framework, uses production NN applications (MLPs, CNNs, and LSTMs) that represent 95% of our datacenters' NN inference demand. Despite low utilization for some applications, the TPU is on average about 15X - 30X faster than its contemporary GPU or CPU, with TOPS/Watt about 30X - 80X higher. Moreover, using the GPU's GDDR5 memory in the TPU would triple achieved TOPS and raise TOPS/Watt to nearly 70X the GPU and 200X the CPU.

In-Datacenter Performance Analysis of a Tensor Processing Unit

Many architects believe that major improvements in cost-energy-performance must now come from domain-specific hardware. This paper evaluates a custom ASIC---called a Tensor Processing Unit (TPU) --- deployed in datacenters since 2015 that accelerates the inference phase of neural networks (NN). The heart of the TPU is a 65,536 8-bit MAC matrix multiply unit that offers a peak throughput of 92 TeraOps/second (TOPS) and a large (28 MiB) software-managed on-chip memory. The TPU's deterministic execution model is a better match to the 99th-percentile response-time requirement of our NN applications than are the time-varying optimizations of CPUs and GPUs that help average throughput more than guaranteed latency. The lack of such features helps explain why, despite having myriad MACs and a big memory, the TPU is relatively small and low power. We compare the TPU to a server-class Intel Haswell CPU and an Nvidia K80 GPU, which are contemporaries deployed in the same datacenters. Our workload, written in the high-level TensorFlow framework, uses production NN applications (MLPs, CNNs, and LSTMs) that represent 95% of our datacenters' NN inference demand. Despite low utilization for some applications, the TPU is on average about 15X -- 30X faster than its contemporary GPU or CPU, with TOPS/Watt about 30X -- 80X higher. Moreover, using the CPU's GDDR5 memory in the TPU would triple achieved TOPS and raise TOPS/Watt to nearly 70X the GPU and 200X the CPU.

/pdf/in-datacenter-performance-analysis-of-a-tensor-processing-4pmp83zqw4.pdf

State-of-the-art deep neural networks (DNNs) have hundreds of millions of connections and are both computationally and memory intensive, making them difficult to deploy on embedded systems with limited hardware resources and power budgets. While custom hardware helps the computation, fetching weights from DRAM is two orders of magnitude more expensive than ALU operations, and dominates the required power.Previously proposed 'Deep Compression' makes it possible to fit large DNNs (AlexNet and VGGNet) fully in on-chip SRAM. This compression is achieved by pruning the redundant connections and having multiple connections share the same weight. We propose an energy efficient inference engine (EIE) that performs inference on this compressed network model and accelerates the resulting sparse matrix-vector multiplication with weight sharing. Going from DRAM to SRAM gives EIE 120× energy saving; Exploiting sparsity saves 10×; Weight sharing gives 8×; Skipping zero activations from ReLU saves another 3×. Evaluated on nine DNN benchmarks, EIE is 189× and 13× faster when compared to CPU and GPU implementations of the same DNN without compression. EIE has a processing power of 102 GOPS working directly on a compressed network, corresponding to 3 TOPS on an uncompressed network, and processes FC layers of AlexNet at 1.88×104 frames/sec with a power dissipation of only 600mW. It is 24,000× and 3,400× more energy efficient than a CPU and GPU respectively. Compared with DaDianNao, EIE has 2.9×, 19× and 3× better throughput, energy efficiency and area efficiency.

EIE: efficient inference engine on compressed deep neural network

Deep neural networks (DNNs) are currently widely used for many artificial intelligence (AI) applications including computer vision, speech recognition, and robotics. While DNNs deliver state-of-the-art accuracy on many AI tasks, it comes at the cost of high computational complexity. Accordingly, techniques that enable efficient processing of DNNs to improve energy efficiency and throughput without sacrificing application accuracy or increasing hardware cost are critical to the wide deployment of DNNs in AI systems. This article aims to provide a comprehensive tutorial and survey about the recent advances toward the goal of enabling efficient processing of DNNs. Specifically, it will provide an overview of DNNs, discuss various hardware platforms and architectures that support DNNs, and highlight key trends in reducing the computation cost of DNNs either solely via hardware design changes or via joint hardware design and DNN algorithm changes. It will also summarize various development resources that enable researchers and practitioners to quickly get started in this field, and highlight important benchmarking metrics and design considerations that should be used for evaluating the rapidly growing number of DNN hardware designs, optionally including algorithmic codesigns, being proposed in academia and industry. The reader will take away the following concepts from this article: understand the key design considerations for DNNs; be able to evaluate different DNN hardware implementations with benchmarks and comparison metrics; understand the tradeoffs between various hardware architectures and platforms; be able to evaluate the utility of various DNN design techniques for efficient processing; and understand recent implementation trends and opportunities.

Efficient Processing of Deep Neural Networks: A Tutorial and Survey

Artificial Neural Networks have dramatically improved performance for many machine learning tasks. We demonstrate a new architecture for a fully optical neural network that enables a computational speed enhancement of at least two orders of magnitude and three orders of magnitude in power efficiency over state-of-the-art electronics.

Deep learning with coherent nanophotonic circuits

A number of recent efforts have attempted to design accelerators for popular machine learning algorithms, such as those involving convolutional and deep neural networks (CNNs and DNNs). These algorithms typically involve a large number of multiply-accumulate (dot-product) operations. A recent project, DaDianNao, adopts a near data processing approach, where a specialized neural functional unit performs all the digital arithmetic operations and receives input weights from adjacent eDRAM banks.This work explores an in-situ processing approach, where memristor crossbar arrays not only store input weights, but are also used to perform dot-product operations in an analog manner. While the use of crossbar memory as an analog dot-product engine is well known, no prior work has designed or characterized a full-fledged accelerator based on crossbars. In particular, our work makes the following contributions: (i) We design a pipelined architecture, with some crossbars dedicated for each neural network layer, and eDRAM buffers that aggregate data between pipeline stages. (ii) We define new data encoding techniques that are amenable to analog computations and that can reduce the high overheads of analog-to-digital conversion (ADC). (iii) We define the many supporting digital components required in an analog CNN accelerator and carry out a design space exploration to identify the best balance of memristor storage/compute, ADCs, and eDRAM storage on a chip. On a suite of CNN and DNN workloads, the proposed ISAAC architecture yields improvements of 14.8×, 5.5×, and 7.5× in throughput, energy, and computational density (respectively), relative to the state-of-the-art DaDianNao architecture.

ISAAC: a convolutional neural network accelerator with in-situ analog arithmetic in crossbars

Historically, server designers have opted for simple memory systems by picking one of a few commoditized DDR memory products. We are already witnessing a major upheaval in the off-chip memory hierarchy, with the introduction of many new memory products—buffer-on-board, LRDIMM, HMC, HBM, and NVMs, to name a few. Given the plethora of choices, it is expected that different vendors will adopt different strategies for their high-capacity memory systems, often deviating from DDR standards and/or integrating new functionality within memory systems. These strategies will likely differ in their choice of interconnect and topology, with a significant fraction of memory energy being dissipated in I/O and data movement. To make the case for memory interconnect specialization, this paper makes three contributions.First, we design a tool that carefully models I/O power in the memory system, explores the design space, and gives the user the ability to define new types of memory interconnects/topologies. The tool is validated against SPICE models, and is integrated into version 7 of the popular CACTI package. Our analysis with the tool shows that several design parameters have a significant impact on I/O power.We then use the tool to help craft novel specialized memory system channels. We introduce a new relay-on-board chip that partitions a DDR channel into multiple cascaded channels. We show that this simple change to the channel topology can improve performance by 22% for DDR DRAM and lower cost by up to 65% for DDR DRAM. This new architecture does not require any changes to DIMMs, and it efficiently supports hybrid DRAM/NVM systems.Finally, as an example of a more disruptive architecture, we design a custom DIMM and parallel bus that moves away from the DDR3/DDR4 standards. To reduce energy and improve performance, the baseline data channel is split into three narrow parallel channels and the on-DIMM interconnects are operated at a lower frequency. In addition, this allows us to design a two-tier error protection strategy that reduces data transfers on the interconnect. This architecture yields a performance improvement of 18% and a memory power reduction of 23%.The cascaded channel and narrow channel architectures serve as case studies for the new tool and show the potential for benefit from re-organizing basic memory interconnects.

https://dl.acm.org/doi/pdf/10.1145/3085572

CACTI 7: New Tools for Interconnect Exploration in Innovative Off-Chip Memories

Intel's SGX offers state-of-the-art security features, including confidentiality, integrity, and authentication (CIA) when accessing sensitive pages in memory. Sensitive pages are placed in an Enclave Page Cache (EPC) within the physical memory before they can be accessed by the processor. To control the overheads imposed by CIA guarantees, the EPC operates with a limited capacity (currently 128 MB). Because of this limited EPC size, sensitive pages must be frequently swapped between EPC and non-EPC regions in memory. A page swap is expensive (about 40K cycles) because it requires an OS system call, page copying, updates to integrity trees and metadata, etc. Our analysis shows that the paging overhead can slow the system on average by 5×, and other studies have reported even higher slowdowns for memory-intensive workloads. The paging overhead can be reduced by growing the size of the EPC to match the size of physical memory, while allowing the EPC to also accommodate non-sensitive pages. However, at least two important problems must be addressed to enable this growth in EPC: (i) the depth of the integrity tree and its cacheability must be improved to keep memory bandwidth overheads in check, (ii) the space overheads of integrity verification (tree and MACs) must be reduced. We achieve both goals by introducing a variable arity unified tree (VAULT) organization that is more compact and has lower depth. We further reduce the space overheads with techniques that combine MAC sharing and compression. With simulations, we show that the combination of our techniques can address most inefficiencies in SGX memory access and improve overall performance by 3.7×, relative to an SGX baseline, while incurring a memory capacity over-head of only 4.7%.

https://dl.acm.org/doi/pdf/10.1145/3296957.3177155

VAULT: Reducing Paging Overheads in SGX with Efficient Integrity Verification Structures

Memory compression has been proposed and deployed in the past to grow the capacity of a memory system and reduce page fault rates. Compression also has secondary benefits: it can reduce energy and bandwidth demands. However, most prior mechanisms have been designed to focus on the capacity metric and few prior works have attempted to explicitly reduce energy or bandwidth. Further, mechanisms that focus on the capacity metric also require complex logic to locate the requested data in memory. In this paper, we design a highly simple compressed memory architecture that does not target the capacity metric. Instead, it focuses on complexity, energy, bandwidth, and reliability. It relies on rank subsetting and a careful placement of compressed data and metadata to achieve these benefits. Further, the space made available via compression is used to boost other metrics - the space can be used to implement stronger error correction codes or energy-efficient data encodings. The best performing MemZip configuration yields a 45% performance improvement and 57% memory energy reduction, compared to an uncompressed non-sub-ranked baseline. Another energy-optimized configuration yields a 29.8% performance improvement and a 79% memory energy reduction, relative to the same baseline.

/pdf/memzip-exploring-unconventional-benefits-from-memory-183384dscv.pdf

MemZip: Exploring unconventional benefits from memory compression

Three-dimensional network-on-chips are suitable communication fabrics for high-density 3D many-core ICs. Such networks have shorter communication hop count, compared to 2D NoCs, and enjoy fast and power efficient TSV wires in vertical links. Unfortunately, the fabrication process of TSV connections has not matured yet, which results in poor vertical links yield. In this work, we address this challenge and introduce AFRA, a deadlock-free routing algorithm for 3D mesh-based NoCs that tolerates faults on vertical links. AFRA is designed to be simple, high performance, and robust. The simplicity is achieved by applying ZXY and XZXY routings in the absence and presence of fault, respectively. Furthermore, AFRA, as will be proved, is deadlock-free when all vertical faulty links have the same direction. This enables the routing to save virtual channels for performance rather than scarifying them for deadlock avoidance. Finally, AFRA provides robustness, which means supporting connection for all possible pairs of communicating nodes in high fault rates. AFRA is evaluated, though cycle accurate network simulation, and is compared with planar adaptive routing. Results reveal that AFRA significantly outperforms planar adaptive routing in both synthetic and real traffic patterns. In addition, the robustness of AFRA is calculated analytically.

Ali Shafiee

Papers

ISAAC: a convolutional neural network accelerator with in-situ analog arithmetic in crossbars

CACTI 7: New Tools for Interconnect Exploration in Innovative Off-Chip Memories

VAULT: Reducing Paging Overheads in SGX with Efficient Integrity Verification Structures

MemZip: Exploring unconventional benefits from memory compression

AFRA: a low cost high performance reliable routing for 3D mesh NoCs