General-purpose code acceleration with limited-precision analog computation
ACM SIGARCH Computer Architecture NewsPublished 14 June 2014
Renée St. Amant, Amir Yazdanbakhsh, Jongse Park, Bradley Thwaites, Hadi Esmaeilzadeh, Arjang Hassibi
Citations148
Generate an AI Snapshot to get a quick, structured summary of this paper.
Study Snapshot
ObjectiveStudy objective
MethodsResearch methodology
PopulationPopulation studied
Sample sizeSample sizes
OutcomesStudy outcomes here
ResultsStudy results comes here
LimitationsResearch study limitations comes here
A concise AI-generated summary of the paper will appear here once you click Generate AI Snapshot.
TL;DR
An algorithmic transformation that automatically converts approximable regions of code from a von Neumann model to an “analog” neural model is utilized that enables general-purpose use of limited-precision, analog hardware to accelerate “approximable” code-code that can tolerate imprecise execution.
Abstract
As improvements in per-transistor speed and energy efficiency diminish, radical departures from conventional approaches are becoming critical to improving the performance and energy efficiency of general-purpose processors. We propose a solution--from circuit to compiler-that enables general-purpose use of limited-precision, analog hardwareto accelerate "approximable" code---code that can tolerate imprecise execution. We utilize an algorithmic transformation that automatically converts approximable regions of code from a von Neumann model to an "analog" neural model. We outline the challenges of taking an analog approach, including restricted-range value encoding, limited precision in computation, circuit inaccuracies, noise, and constraints on supported topologies. We address these limitations with a combination of circuit techniques, a hardware/software interface, neuralnetwork training techniques, and compiler support. Analog neural acceleration provides whole application speedup of 3.7x and energy savings of 6.3x with quality loss less than 10% for all except one benchmark. These results show that using limited-precision analog circuits for code acceleration, through a neural approach, is both feasible and beneficial over a range of approximation-tolerant, emerging applications including financial analysis, signal processing, robotics, 3D gaming, compression, and image processing
Keywords
Computer ScienceEngineering
Learning Internal Representations by Error Propagation
16,235 Citations1985David E. Rumelhart, Geoffrey E. Hinton +1 more
IntegrationCMOS analog circuit design
3,214 Citations1988
This chapter discusses the design of two-STAGE, OPEN-LOOP COMPARATOR DESIGN, and improving the performance of open-loop COMPARATORS, as well as other aspects of CMOS technology.
Elsevier eBooksLearning Internal Representations by Error Propagation
3,013 Citations1988David E. Rumelhart, Geoffrey E. Hinton +1 more
This chapter contains sections titled: The Problem, The Generalized Delta Rule, Simulation Results, Some Further Generalizations, Conclusion.
Analog Integrated Circuit Design
2,516 Citations1996Anthony Chan Carusone, D.A. Johns +1 more
McPAT
2,342 Citations2009Sheng Li, Jung Ho Ahn +4 more
Combining power, area, and timing results of McPAT with performance simulation of PARSEC benchmarks at the 22nm technology node for both common in-order and out-of-order manycore designs shows that when die cost is not taken into account clustering 8 cores together gives the best energy-delay product, whereas when cost is taking into account configuring clusters with 4 cores gives thebest EDA2P and EDAP.
Dark silicon and the end of multicore scaling
1,511 Citations2011Hadi Esmaeilzadeh, Emily Blem +3 more
The study shows that regardless of chip organization and topology, multicore scaling is power limited to a degree not widely appreciated by the computing community.
Dynamic Warp Formation and Scheduling for Efficient GPU Control Flow
727 Citations2007Wilson Wai Lun Fung, Ivan Sham +2 more
This work implements two major extensions to the CACTI cache modeling tool that focus on interconnect design for a large cache, and adopts state-of-the-art design space exploration strategies for non-uniform cache access (NUCA).
Neural Acceleration for General-Purpose Approximate Programs
618 Citations2012Hadi Esmaeilzadeh, Adrian Sampson +2 more
A programming model is defined that allows programmers to identify approximable code regions -- code that can produce imprecise but acceptable results and offloading approximable code regions to NPUs is faster and more energy efficient than executing the original code.
Optimizing NUCA Organizations and Wiring Alternatives for Large Caches with CACTI 6.0
482 Citations2007Xiaoyao Liang, Ramón Canal +1 more
A range of cache refresh and placement schemes that are sensitive to retention time are proposed, and it is shown that most of the retention time variations can be masked by the microarchitecture when using these schemes.
Green
435 Citations2010Woongki Baek, Trishul Chilimbi
Green enables programmers to approximate expensive functions and loops and operates in two phases and can produce significant improvements in performance and energy consumption with small and controlled QoS degradation.
Flikker
409 Citations2011Song Liu, Karthik Pattabiraman +2 more
Flikker exposes and leverages an interesting trade-off between energy consumption and hardware correctness, and shows that many applications are naturally tolerant to errors in the non-critical data, and in the vast majority of cases, the errors have little or no impact on the application's final outcome.
Architecture support for disciplined approximate programming
385 Citations2012Hadi Esmaeilzadeh, Adrian Sampson +2 more
An ISA extension that provides approximate operations and storage is described that gives the hardware freedom to save energy at the cost of accuracy and Truffle, a microarchitecture design that efficiently supports the ISA extensions is proposed.
EnerJ
384 Citations2011Adrian Sampson, Werner Dietl +4 more
EnerJ is developed, an extension to Java that adds approximate data types and a hardware architecture that offers explicit approximate storage and computation and allows a programmer to control explicitly how information flows from approximate data to precise data.
A high-performance microarchitecture with hardware-programmable functional units
359 Citations1994Rahul Razdan, Michael D. Smith
A novel way to incorporate hardware-programmable resources into a processor microarchitecture to improve the performance of general-purpose applications through a coupling of compile-time analysis routines and hardware synthesis tools is explored.
MARSS
359 Citations2011Avadh Patel, Furat Afram +2 more
MARSS simulates the execution of all software components in the system, including unmodified binaries of applications, OS and libraries, as well as detailed models of coherent caches, interconnections, chipsets, memory and IO devices.
Improving the Rprop Learning Algorithm
350 Citations2000Christian Igel, Michael Hüsken
Modifications of the Rprop algorithm are introduced that improve its learning speed and the resulting speedup is experimentally shown for a set of neural network learning tasks as well as for artificial error surfaces.
IEEE MicroToward Dark Silicon in Servers
311 Citations2011Nikos Hardavellas, Michael Ferdman +2 more
Server chips will not scale beyond a few tens to low hundreds of cores, and an increasing fraction of the chip in future technologies will be dark silicon that the authors cannot afford to power.
Conservation cores
273 Citations2010Ganesh Venkatesh, Jack Sampson +6 more
A toolchain for automatically synthesizing c-cores from application source code is presented and it is demonstrated that they can significantly reduce energy and energy-delay for a wide range of applications, and patching can extend the useful lifetime of individual c-Cores to match that of conventional processors.
Wafer-scale integration of analog neural networks
272 Citations2008Johannes Schemmel, Johannes Fieres +1 more
A novel asynchronous low-voltage signaling scheme is presented that makes the wafer-scale approach feasible by limiting the total power consumption while simultaneously providing a flexible, programmable network topology.
SAGE
251 Citations2013Mehrzad Samadi, Janghaeng Lee +3 more
Across a set of machine learning and image processing kernels, SAGE's approximation yields an average of 2.5× speedup with less than 10% quality loss compared to the accurate execution on a NVIDIA GTX 560 GPU.
Quality programmable vector processors for approximate computing
250 Citations2013Swagath Venkataramani, Vinay K. Chippa +3 more
Quality programmable processors, in which the notion of quality is explicitly codified in the HW/SW interface, are suggested to be a significant step towards bringing approximate computing to the mainstream.
Verifying quantitative reliability for programs that execute on unreliable hardware
223 Citations2013Michael Carbin, Saša Misailovíc +1 more
This work presents Rely a programming language that enables developers to reason about the quantitative reliability of an application -- namely, the probability that it produces the correct result when executed on unreliable hardware.
Quality of service profiling
221 Citations2010Saša Misailovíc, Stelios Sidiroglou +2 more
The experimental results from applying the implemented quality of service profiler to a challenging set of benchmark applications show that it can enable developers to identify promising optimization opportunities and deliver successful optimizations that substantially increase the performance with only smallquality of service losses.
Approximate storage in solid-state memories
213 Citations2013Adrian Sampson, Jacob Nelson +2 more
Simulations show that reduced-precision writes in multi-level phase-change memory cells can be 1.7× faster on average and using failed blocks can improve array lifetime by 23% on average with quality loss under 10%.
Energy-efficient signal processing via algorithmic noise-tolerance
207 Citations1999Rajamohana Hegde, Naresh R. Shanbhag
A framework for low-energy digital signal processing (DSP) where the supply voltage is scaled beyond the critical voltage required to match the critical path delay to the throughput is proposed and a prediction based error-control scheme is proposed to enhance the performance of the filtering algorithm in presence of errors due to soft computations.
Dynamically Specialized Datapaths for energy efficient computing
204 Citations2011Venkatraman Govindaraju, Chen-Han Ho +1 more
D Dynamically Specialized Datapaths are proposed to improve the energy efficiency of general purpose programmable processors and show that in most cases two DySER blocks can achieve the same performance as having a specialized hardware module for each path-tree.
Relax
201 Citations2010Marc de Kruijf, Shuou Nomura +1 more
This paper considers whether exposing hardware fault information to software and allowing software to control fault recovery simplifies hardware design and helps technology scaling, and describes Relax, an architectural framework for software recovery of hardware faults.
Application-Specific Processing on a General-Purpose Core via Transparent Instruction Set Customization
199 Citations2005Nathan Clark, Manjunath Kudlur +3 more
This work proposes a strategy to transparent customization of the core computation capabilities of the processor without changing its instruction set, and performs subgraph identification at run-time, replacing them with new microcode instructions to configure and utilize the array.
Lecture notes in computer scienceFPGA Implementations of Neural Networks – A Survey of a Decade of Progress
196 Citations2003Jihan Zhu, Peter Sutton
A taxonomy for classifying FPGA implementations of artificial neural networks (ANNs) is provided and different implementation techniques and design issues are discussed.
Hardware spiking neurons design: Analog or digital?
184 Citations2012Antoine Joubert, Bilel Belhadj +2 more
This paper compares the digital and analog implementations of the same Leaky Integrate-and-Fire neuron model at the same technology node with the same level of performance and energy, and shows that the analog implementation requires 5 times less area, and consumes 20 times less energy than the digital design.
IEEE Transactions on ComputersFuzzy Memoization for Floating-Point Multimedia Applications
151 Citations2005Carlos Álvarez, Jesús Corbal +1 more
A novel way of understanding multimedia floating-point operations based on the fuzzy computation paradigm: performance and power consumption can be improved at the cost of small precision losses in computation.
ACM SIGARCH Computer Architecture NewsA defect-tolerant accelerator for emerging high-performance applications
141 Citations2012Olivier Temam
It is empirically show that the conceptual error tolerance of neural networks does translate into the defect tolerance of hardware neural networks, paving the way for their introduction in heterogeneous multi-cores as intrinsically defect-tolerant and energy-efficient accelerators.
IEEE Journal of Solid-State CircuitsAn analog neural network processor with programmable topology
139 Citations1991Bernhard E. Boser, Eduard Säckinger +3 more
The architecture, implementation, and applications of a special-purpose neural network processor are described and the practicality of the chip is demonstrated with an implementation of a neural network for optical character recognition.
IEEE Transactions on ComputersEnergy-Efficient Floating-Point Unit Design
132 Citations2010Sameh Galal, Mark Horowitz
This work presents a method for creating a trade-off curve that can be used to estimate the maximum floating-point performance given a set of area and power constraints, and finds that in a 90 nm CMOS technology at 1 W/mm2, one can achieve a performance of 27 GFlops/ mm2 single precision, and 7.5 GFlop/mm double precision.
Bundled execution of recurring traces for energy-efficient general purpose processing
112 Citations2011Shantanu Gupta, Shuguang Feng +3 more
This approach identifies recurring instruction sequences as phases of “temporal regularity” in a program's execution, and maps suitable ones to the BERET hardware, a three-stage pipeline with a bundled execution model that demonstrates significant savings on instruction fetch, decode and register file accesses energy.
Scalable stochastic processors
110 Citations2010Sriram Narayanan, John Sartori +2 more
BenchNN: On the broad potential application scope of hardware neural network accelerators
109 Citations2012Tianshi Chen, Yunji Chen +8 more
Software neural network implementations of 5 RMS applications from the PARSEC Benchmark Suite are developed and evaluated and it is highlighted that a hardware neural network accelerator is indeed compatible with many of the emerging high- performance workloads, currently accepted as benchmarks for high-performance micro-architectures.
Ultra-Efficient (Embedded) SOC Architectures based on Probabilistic CMOS (PCMOS) Technology
108 Citations2006Lakshmi Narasimhan Chakrapani, B.E.S. Akgul +4 more
This paper shows that PCMOS technology yields significant improvements, both in the energy consumed as well as in the performance, for probabilistic applications with broad utility, using an application-architecture-technology (A2T) co-design methodology introduced here.
QsCores
108 Citations2011Ganesh Venkatesh, Jack Sampson +4 more
While the transistor budget continues to increase exponentially, the power budget has become the dominant limiting factor in processor design and utilizing transistors to design specialized cores that optimize energy-per-computation becomes an effective approach to improve system performance.
Neural NetworksOn the capabilities of neural networks using limited precision weights
93 Citations2002Sorin Drăghici
The paper presents an existence result which relates the difficulty of the problem as characterized by the minimum distance between patterns of different classes to the weight range necessary to ensure that a solution exists, which allows us to calculate a weight range for a given category of problems and be confident that the network has the capability to solve the given problems with integer weights in that range.
Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIEWeight discretization paradigm for optical neural networks
89 Citations1990Emile Fiesler, A. Choudry +1 more
In this paper a weight discretization paradigm is presented for back(ward error) propagation neural networks which can work with a very limited number of discretized levels.
A defect-tolerant accelerator for emerging high-performance applications
89 Citations2012Olivier Temam
Bridging the computation gap between programmable processors and hardwired accelerators
76 Citations2009Kevin Fan, Manjunath Kudlur +2 more
This paper proposes a customized semi-programmable loop accelerator architecture that exploits the efficiency gains available through high levels of customization, while maintaining sufficient flexibility to execute multiple similar loops.
Elastic CGRAs
69 Citations2013Yuanjie Huang, Paolo Ienne +3 more
Elastic CGRAs are introduced, the superscalar processors of computing fabrics: no complex schedule needs to be computed at configuration time, and the operations execute dynamically in the CGRA when data are ready, thus exploiting the data parallelism that an application offers.
Continuous real-world inputs can open up alternative accelerator designs
64 Citations2013Bilel Belhadj, Antoine Joubert +3 more
This article introduces a configurable hybrid digital/analog micro-architecture capable of implementing a broad set of signal processing applications based on these concepts, together with a back-end optimizer which takes advantage of the special nature of these applications.
Automatic abstraction and fault tolerance in cortical microachitectures
54 Citations2011Atif Hashmi, Hugues Berry +2 more
A stuck-at fault model for cortical networks that mimic the human cortex is introduced, and it is shown experimentally that the model software implementation can intrinsically preserve its functionality in the presence of faulty hardware, without requiring any reprogramming or recompilation.
CHiMPS
47 Citations2008Andrew Putnam, Dave Bennett +3 more
CHiMPS is a C-based accelerator compiler for hybrid CPU-FPGA computing platforms that inputs generic ANSIC code and automatically generates VHDL blocks for an FPGA.
Parallel digital implementations of neural networks
45 Citations2002K. Wojtek Przytula
The paper reviews implementations of neural networks on parallel digital machines and the connectionist neural networks models are discussed from the point of view of their computational characteristics.
A case for neuromorphic ISAs
40 Citations2011Atif Hashmi, Andrew Nere +2 more
A well-defined abstraction layer -- the Neuromorphic instruction set architecture, or NISA -- is proposed that separates a neural application's algorithmic specification from the underlying execution substrate, and the Aivo framework is described, which demonstrates the concrete advantages of such an abstraction layer.
Learning on an analog VLSI neural network chip
38 Citations2002Simon Tam, Bharat Gupta +2 more
The issues associated with implementing the error backpropagation algorithm on a 64-neuron nonvolatile analog VLSI neural network chip (ETANN) are described and an adaptive, reinforced, bake-train-bake scheme is reported.
An Analog Neural Network Processor With Programmable Network Topology
35 Citations1991Bernhard E. Boser, Eduard Säckinger
A Fault Criticality Evaluation Framework of Digital Systems for Error Tolerant Video Applications
34 Citations2011Yuntan Fang, Huawei Li +1 more
A framework across multiple layers for fault criticality evaluation is proposed, taking an H.264/AVC decoder as an example, to demonstrate that for different functional modules, the faults in them bear different fault criticalities because of their unbalanced effects on applications.
Exploiting Soft Computing for Increased Fault Tolerance
34 Citations2006Xuanhua Li, Donald Yeung
This paper identifies three characteristics of soft computations that make them resilient to error: redundancy, adaptivity, and reduced precision, and presents a method for identifyingsoft computations at the instruction level using dynamic slicing analysis.
Neural Network Stream Processing Core (NnSP) for Embedded Systems
34 Citations2006Hadi Esmaeilzadeh, Parisa Saeedi +3 more
The obtained results show that a 64-PE NnSP can perform computations of 25.6 giga connections in a second, while its throughput is upto 51.2 giga 32-bit fixed point operations per second, placing it among the best state of the art parallel processors.
IEEE MicroMixed-Signal Approximate Computation: A Neural Predictor Case Study
11 Citations2009Renée St. Amant, Daniel A. Jiménez +1 more
An enhanced neural prediction algorithm and its efficient analog implementation is described, which allows similar computation to be performed at higher speed using less power.