Approximate computing is an emerging paradigm enabling tradeoffs between accuracy and efficiency. However, a fundamental challenge persists: state-of-the-art techniques lack the ability to enforce runtime guarantees on accuracy. The convention is to 1) employ offline or online accuracy models, or 2) present experimental results that demonstrate empirically low error. Unfortunately, these approaches are still unable to guarantee acceptability of all application outputs at runtime. We offer a solution that revisits concepts from anytime algorithms. Originally explored for real-time decision problems, anytime algorithms have the property of producing results with increasing accuracy over time. We propose the Anytime Automaton, a new computation model that executes applications as a parallel pipeline of anytime approximations. An automaton produces approximate versions of the application output with increasing accuracy, guaranteeing that the final precise version is eventually reached. The automaton can be stopped whenever the output is deemed acceptable, otherwise, it is a simple matter of letting it run longer. We present an in-depth analysis of the model and demonstrate attractive runtime-accuracy profiles on various applications. Our anytime automaton is the first step towards systems where the acceptability of an application's output directly governs the amount of time and energy expended.
{"title":"The Anytime Automaton","authors":"Joshua San Miguel, Natalie D. Enright Jerger","doi":"10.1145/3007787.3001195","DOIUrl":"https://doi.org/10.1145/3007787.3001195","url":null,"abstract":"Approximate computing is an emerging paradigm enabling tradeoffs between accuracy and efficiency. However, a fundamental challenge persists: state-of-the-art techniques lack the ability to enforce runtime guarantees on accuracy. The convention is to 1) employ offline or online accuracy models, or 2) present experimental results that demonstrate empirically low error. Unfortunately, these approaches are still unable to guarantee acceptability of all application outputs at runtime. We offer a solution that revisits concepts from anytime algorithms. Originally explored for real-time decision problems, anytime algorithms have the property of producing results with increasing accuracy over time. We propose the Anytime Automaton, a new computation model that executes applications as a parallel pipeline of anytime approximations. An automaton produces approximate versions of the application output with increasing accuracy, guaranteeing that the final precise version is eventually reached. The automaton can be stopped whenever the output is deemed acceptable, otherwise, it is a simple matter of letting it run longer. We present an in-depth analysis of the model and demonstrate attractive runtime-accuracy profiles on various applications. Our anytime automaton is the first step towards systems where the acceptability of an application's output directly governs the amount of time and energy expended.","PeriodicalId":6634,"journal":{"name":"2016 ACM/IEEE 43rd Annual International Symposium on Computer Architecture (ISCA)","volume":"782 1","pages":"545-557"},"PeriodicalIF":0.0,"publicationDate":"2016-06-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"89229621","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Deep convolutional neural networks (CNNs) are widely used in modern AI systems for their superior accuracy but at the cost of high computational complexity. The complexity comes from the need to simultaneously process hundreds of filters and channels in the high-dimensional convolutions, which involve a significant amount of data movement. Although highly-parallel compute paradigms, such as SIMD/SIMT, effectively address the computation requirement to achieve high throughput, energy consumption still remains high as data movement can be more expensive than computation. Accordingly, finding a dataflow that supports parallel processing with minimal data movement cost is crucial to achieving energy-efficient CNN processing without compromising accuracy. In this paper, we present a novel dataflow, called row-stationary (RS), that minimizes data movement energy consumption on a spatial architecture. This is realized by exploiting local data reuse of filter weights and feature map pixels, i.e., activations, in the high-dimensional convolutions, and minimizing data movement of partial sum accumulations. Unlike dataflows used in existing designs, which only reduce certain types of data movement, the proposed RS dataflow can adapt to different CNN shape configurations and reduces all types of data movement through maximally utilizing the processing engine (PE) local storage, direct inter-PE communication and spatial parallelism. To evaluate the energy efficiency of the different dataflows, we propose an analysis framework that compares energy cost under the same hardware area and processing parallelism constraints. Experiments using the CNN configurations of AlexNet show that the proposed RS dataflow is more energy efficient than existing dataflows in both convolutional (1.4× to 2.5×) and fully-connected layers (at least 1.3× for batch size larger than 16). The RS dataflow has also been demonstrated on a fabricated chip, which verifies our energy analysis.
{"title":"Eyeriss: A Spatial Architecture for Energy-Efficient Dataflow for Convolutional Neural Networks","authors":"Yu-hsin Chen, J. Emer, V. Sze","doi":"10.1145/3007787.3001177","DOIUrl":"https://doi.org/10.1145/3007787.3001177","url":null,"abstract":"Deep convolutional neural networks (CNNs) are widely used in modern AI systems for their superior accuracy but at the cost of high computational complexity. The complexity comes from the need to simultaneously process hundreds of filters and channels in the high-dimensional convolutions, which involve a significant amount of data movement. Although highly-parallel compute paradigms, such as SIMD/SIMT, effectively address the computation requirement to achieve high throughput, energy consumption still remains high as data movement can be more expensive than computation. Accordingly, finding a dataflow that supports parallel processing with minimal data movement cost is crucial to achieving energy-efficient CNN processing without compromising accuracy. In this paper, we present a novel dataflow, called row-stationary (RS), that minimizes data movement energy consumption on a spatial architecture. This is realized by exploiting local data reuse of filter weights and feature map pixels, i.e., activations, in the high-dimensional convolutions, and minimizing data movement of partial sum accumulations. Unlike dataflows used in existing designs, which only reduce certain types of data movement, the proposed RS dataflow can adapt to different CNN shape configurations and reduces all types of data movement through maximally utilizing the processing engine (PE) local storage, direct inter-PE communication and spatial parallelism. To evaluate the energy efficiency of the different dataflows, we propose an analysis framework that compares energy cost under the same hardware area and processing parallelism constraints. Experiments using the CNN configurations of AlexNet show that the proposed RS dataflow is more energy efficient than existing dataflows in both convolutional (1.4× to 2.5×) and fully-connected layers (at least 1.3× for batch size larger than 16). The RS dataflow has also been demonstrated on a fabricated chip, which verifies our energy analysis.","PeriodicalId":6634,"journal":{"name":"2016 ACM/IEEE 43rd Annual International Symposium on Computer Architecture (ISCA)","volume":"28 1","pages":"367-379"},"PeriodicalIF":0.0,"publicationDate":"2016-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"73822920","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
With the degree of parallelism increasing, performance of multi-threaded shared variable applications is not only limited by serialized critical section execution, but also by the serialized competition overhead for threads to get access to critical section. As the number of concurrent threads grows, such competition overhead may exceed the time spent in critical section itself, and become the dominating factor limiting the performance of parallel applications. In modern operating systems, queue spinlock, which comprises a low-overhead spinning phase and a high-overhead sleeping phase, is often used to lock critical sections. In the paper, we show that this advanced locking solution may create very high competition overhead for multithreaded applications executing in NoC-based CMPs. Then we propose a software-hardware cooperative mechanism that can opportunistically maximize the chance that a thread wins the critical section access in the low-overhead spinning phase, thereby reducing the competition overhead. At the OS primitives level, we monitor the remaining times of retry (RTR) in a thread's spinning phase, which reflects in how long the thread must enter into the high-overhead sleep mode. At the hardware level, we integrate the RTR information into the packets of locking requests, and let the NoC prioritize locking request packets according to the RTR information. The principle is that the smaller RTR a locking request packet carries, the higher priority it gets and thus quicker delivery. We evaluate our opportunistic competition overhead reduction technique with cycle-accurate full-system simulations in GEM5 using PARSEC (11 programs) and SPEC OMP2012 (14 programs) benchmarks. Compared to the original queue spinlock implementation, experimental results show that our method can effectively increase the opportunity of threads entering the critical section in low-overhead spinning phase, reducing the competition overhead averagely by 39.9% (maximally by 61.8%) and accelerating the execution of the Region-of-Interest averagely by 14.4% (maximally by 24.5%) across all 25 benchmark programs.
{"title":"Opportunistic Competition Overhead Reduction for Expediting Critical Section in NoC Based CMPs","authors":"Y. Yao, Zhonghai Lu","doi":"10.1145/3007787.3001167","DOIUrl":"https://doi.org/10.1145/3007787.3001167","url":null,"abstract":"With the degree of parallelism increasing, performance of multi-threaded shared variable applications is not only limited by serialized critical section execution, but also by the serialized competition overhead for threads to get access to critical section. As the number of concurrent threads grows, such competition overhead may exceed the time spent in critical section itself, and become the dominating factor limiting the performance of parallel applications. In modern operating systems, queue spinlock, which comprises a low-overhead spinning phase and a high-overhead sleeping phase, is often used to lock critical sections. In the paper, we show that this advanced locking solution may create very high competition overhead for multithreaded applications executing in NoC-based CMPs. Then we propose a software-hardware cooperative mechanism that can opportunistically maximize the chance that a thread wins the critical section access in the low-overhead spinning phase, thereby reducing the competition overhead. At the OS primitives level, we monitor the remaining times of retry (RTR) in a thread's spinning phase, which reflects in how long the thread must enter into the high-overhead sleep mode. At the hardware level, we integrate the RTR information into the packets of locking requests, and let the NoC prioritize locking request packets according to the RTR information. The principle is that the smaller RTR a locking request packet carries, the higher priority it gets and thus quicker delivery. We evaluate our opportunistic competition overhead reduction technique with cycle-accurate full-system simulations in GEM5 using PARSEC (11 programs) and SPEC OMP2012 (14 programs) benchmarks. Compared to the original queue spinlock implementation, experimental results show that our method can effectively increase the opportunity of threads entering the critical section in low-overhead spinning phase, reducing the competition overhead averagely by 39.9% (maximally by 61.8%) and accelerating the execution of the Region-of-Interest averagely by 14.4% (maximally by 24.5%) across all 25 benchmark programs.","PeriodicalId":6634,"journal":{"name":"2016 ACM/IEEE 43rd Annual International Symposium on Computer Architecture (ISCA)","volume":"83 1","pages":"279-290"},"PeriodicalIF":0.0,"publicationDate":"2016-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"80337153","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Ali Shafiee, Anirban Nag, Naveen Muralimanohar, R. Balasubramonian, J. Strachan, Miao Hu, R. S. Williams, Vivek Srikumar
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.
{"title":"ISAAC: A Convolutional Neural Network Accelerator with In-Situ Analog Arithmetic in Crossbars","authors":"Ali Shafiee, Anirban Nag, Naveen Muralimanohar, R. Balasubramonian, J. Strachan, Miao Hu, R. S. Williams, Vivek Srikumar","doi":"10.1145/3007787.3001139","DOIUrl":"https://doi.org/10.1145/3007787.3001139","url":null,"abstract":"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.","PeriodicalId":6634,"journal":{"name":"2016 ACM/IEEE 43rd Annual International Symposium on Computer Architecture (ISCA)","volume":"89 1","pages":"14-26"},"PeriodicalIF":0.0,"publicationDate":"2016-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"79400343","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Song Han, Xingyu Liu, Huizi Mao, Jing Pu, A. Pedram, M. Horowitz, W. Dally
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 120x energy saving, Exploiting sparsity saves 10x, Weight sharing gives 8x, Skipping zero activations from ReLU saves another 3x. Evaluated on nine DNN benchmarks, EIE is 189x and 13x 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 1.88x104 frames/sec with a power dissipation of only 600mW. It is 24,000x and 3,400x more energy efficient than a CPU and GPU respectively. Compared with DaDianNao, EIE has 2.9x, 19x and 3x better throughput, energy efficiency and area efficiency.
{"title":"EIE: Efficient Inference Engine on Compressed Deep Neural Network","authors":"Song Han, Xingyu Liu, Huizi Mao, Jing Pu, A. Pedram, M. Horowitz, W. Dally","doi":"10.1145/3007787.3001163","DOIUrl":"https://doi.org/10.1145/3007787.3001163","url":null,"abstract":"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 120x energy saving, Exploiting sparsity saves 10x, Weight sharing gives 8x, Skipping zero activations from ReLU saves another 3x. Evaluated on nine DNN benchmarks, EIE is 189x and 13x 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 1.88x104 frames/sec with a power dissipation of only 600mW. It is 24,000x and 3,400x more energy efficient than a CPU and GPU respectively. Compared with DaDianNao, EIE has 2.9x, 19x and 3x better throughput, energy efficiency and area efficiency.","PeriodicalId":6634,"journal":{"name":"2016 ACM/IEEE 43rd Annual International Symposium on Computer Architecture (ISCA)","volume":"232 1","pages":"243-254"},"PeriodicalIF":0.0,"publicationDate":"2016-02-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"75271239","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Conventional translation look-aside buffers(TLBs) are required to complete address translation withshort latencies, as the address translation is on the criticalpath of all memory accesses even for L1 cache hits. Such strictTLB latency restrictions limit the TLB capacity, as the latencyincrease with large TLBs may lower the overall performanceeven with potential TLB miss reductions. Furthermore, TLBsconsume a significant amount of energy as they are accessedfor every instruction fetch and data access. To avoid thelatency restriction and reduce the energy consumption, virtualcaching techniques have been proposed to defer translation toafter L1 cache misses. However, an efficient solution for thesynonym problem has been a critical issue hindering the wideadoption of virtual caching.Based on the virtual caching concept, this study proposes ahybrid virtual memory architecture extending virtual cachingto the entire cache hierarchy, aiming to improve both performanceand energy consumption. The hybrid virtual cachinguses virtual addresses augmented with address space identifiers(ASID) in the cache hierarchy for common non-synonymaddresses. For such non-synonyms, the address translationoccurs only after last-level cache (LLC) misses. For uncommonsynonym addresses, the addresses are translated to physicaladdresses with conventional TLBs before L1 cache accesses. Tosupport such hybrid translation, we propose an efficient synonymdetection mechanism based on Bloom filters which canidentify synonym candidates with few false positives. For largememory applications, delayed translation alone cannot solvethe address translation problem, as fixed-granularity delayedTLBs may not scale with the increasing memory requirements.To mitigate the translation scalability problem, this studyproposes a delayed many segment translation designed for thehybrid virtual caching. The experimental results show that ourapproach effectively lowers accesses to the TLBs, leading tosignificant power savings. In addition, the approach providesperformance improvement with scalable delayed translationwith variable length segments.
{"title":"Efficient synonym filtering and scalable delayed translation for hybrid virtual caching","authors":"Chang Hyun Park, Taekyung Heo, Jaehyuk Huh","doi":"10.1145/3007787.3001160","DOIUrl":"https://doi.org/10.1145/3007787.3001160","url":null,"abstract":"Conventional translation look-aside buffers(TLBs) are required to complete address translation withshort latencies, as the address translation is on the criticalpath of all memory accesses even for L1 cache hits. Such strictTLB latency restrictions limit the TLB capacity, as the latencyincrease with large TLBs may lower the overall performanceeven with potential TLB miss reductions. Furthermore, TLBsconsume a significant amount of energy as they are accessedfor every instruction fetch and data access. To avoid thelatency restriction and reduce the energy consumption, virtualcaching techniques have been proposed to defer translation toafter L1 cache misses. However, an efficient solution for thesynonym problem has been a critical issue hindering the wideadoption of virtual caching.Based on the virtual caching concept, this study proposes ahybrid virtual memory architecture extending virtual cachingto the entire cache hierarchy, aiming to improve both performanceand energy consumption. The hybrid virtual cachinguses virtual addresses augmented with address space identifiers(ASID) in the cache hierarchy for common non-synonymaddresses. For such non-synonyms, the address translationoccurs only after last-level cache (LLC) misses. For uncommonsynonym addresses, the addresses are translated to physicaladdresses with conventional TLBs before L1 cache accesses. Tosupport such hybrid translation, we propose an efficient synonymdetection mechanism based on Bloom filters which canidentify synonym candidates with few false positives. For largememory applications, delayed translation alone cannot solvethe address translation problem, as fixed-granularity delayedTLBs may not scale with the increasing memory requirements.To mitigate the translation scalability problem, this studyproposes a delayed many segment translation designed for thehybrid virtual caching. The experimental results show that ourapproach effectively lowers accesses to the TLBs, leading tosignificant power savings. In addition, the approach providesperformance improvement with scalable delayed translationwith variable length segments.","PeriodicalId":6634,"journal":{"name":"2016 ACM/IEEE 43rd Annual International Symposium on Computer Architecture (ISCA)","volume":"97 1","pages":"217-229"},"PeriodicalIF":0.0,"publicationDate":"2016-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"90632704","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：