Matching Items (17)
Filtering by
- All Subjects: Electrical Engineering
Description
With increasing transistor volume and reducing feature size, it has become a major design constraint to reduce power consumption also. This has given rise to aggressive architectural changes for on-chip power management and rapid development to energy efficient hardware accelerators. Accordingly, the objective of this research work is to facilitate software developers to leverage these hardware techniques and improve energy efficiency of the system. To achieve this, I propose two solutions for Linux kernel: Optimal use of these architectural enhancements to achieve greater energy efficiency requires accurate modeling of processor power consumption. Though there are many models available in literature to model processor power consumption, there is a lack of such models to capture power consumption at the task-level. Task-level energy models are a requirement for an operating system (OS) to perform real-time power management as OS time multiplexes tasks to enable sharing of hardware resources. I propose a detailed design methodology for constructing an architecture agnostic task-level power model and incorporating it into a modern operating system to build an online task-level power profiler. The profiler is implemented inside the latest Linux kernel and validated for Intel Sandy Bridge processor. It has a negligible overhead of less than 1\% hardware resource consumption. The profiler power prediction was demonstrated for various application benchmarks from SPEC to PARSEC with less than 4\% error. I also demonstrate the importance of the proposed profiler for emerging architectural techniques through use case scenarios, which include heterogeneous computing and fine grained per-core DVFS. Along with architectural enhancement in general purpose processors to improve energy efficiency, hardware accelerators like Coarse Grain reconfigurable architecture (CGRA) are gaining popularity. Unlike vector processors, which rely on data parallelism, CGRA can provide greater flexibility and compiler level control making it more suitable for present SoC environment. To provide streamline development environment for CGRA, I propose a flexible framework in Linux to do design space exploration for CGRA. With accurate and flexible hardware models, fine grained integration with accurate architectural simulator, and Linux memory management and DMA support, a user can carry out limitless experiments on CGRA in full system environment.
ContributorsDesai, Digant Pareshkumar (Author) / Vrudhula, Sarma (Thesis advisor) / Chakrabarti, Chaitali (Committee member) / Wu, Carole-Jean (Committee member) / Arizona State University (Publisher)
Created2013
Description
In recent years we have witnessed a shift towards multi-processor system-on-chips (MPSoCs) to address the demands of embedded devices (such as cell phones, GPS devices, luxury car features, etc.). Highly optimized MPSoCs are well-suited to tackle the complex application demands desired by the end user customer. These MPSoCs incorporate a constellation of heterogeneous processing elements (PEs) (general purpose PEs and application-specific integrated circuits (ASICS)). A typical MPSoC will be composed of a application processor, such as an ARM Coretex-A9 with cache coherent memory hierarchy, and several application sub-systems. Each of these sub-systems are composed of highly optimized instruction processors, graphics/DSP processors, and custom hardware accelerators. Typically, these sub-systems utilize scratchpad memories (SPM) rather than support cache coherency. The overall architecture is an integration of the various sub-systems through a high bandwidth system-level interconnect (such as a Network-on-Chip (NoC)). The shift to MPSoCs has been fueled by three major factors: demand for high performance, the use of component libraries, and short design turn around time. As customers continue to desire more and more complex applications on their embedded devices the performance demand for these devices continues to increase. Designers have turned to using MPSoCs to address this demand. By using pre-made IP libraries designers can quickly piece together a MPSoC that will meet the application demands of the end user with minimal time spent designing new hardware. Additionally, the use of MPSoCs allows designers to generate new devices very quickly and thus reducing the time to market. In this work, a complete MPSoC synthesis design flow is presented. We first present a technique \cite{leary1_intro} to address the synthesis of the interconnect architecture (particularly Network-on-Chip (NoC)). We then address the synthesis of the memory architecture of a MPSoC sub-system \cite{leary2_intro}. Lastly, we present a co-synthesis technique to generate the functional and memory architectures simultaneously. The validity and quality of each synthesis technique is demonstrated through extensive experimentation.
ContributorsLeary, Glenn (Author) / Chatha, Karamvir S (Thesis advisor) / Vrudhula, Sarma (Committee member) / Shrivastava, Aviral (Committee member) / Beraha, Rudy (Committee member) / Arizona State University (Publisher)
Created2013
Description
Multicore processors have proliferated in nearly all forms of computing, from servers, desktop, to smartphones. The primary reason for this large adoption of multicore processors is due to its ability to overcome the power-wall by providing higher performance at a lower power consumption rate. With multi-cores, there is increased need for dynamic energy management (DEM), much more than for single-core processors, as DEM for multi-cores is no more a mechanism just to ensure that a processor is kept under specified temperature limits, but also a set of techniques that manage various processor controls like dynamic voltage and frequency scaling (DVFS), task migration, fan speed, etc. to achieve a stated objective. The objectives span a wide range from maximizing throughput, minimizing power consumption, reducing peak temperature, maximizing energy efficiency, maximizing processor reliability, and so on, along with much more wider constraints of temperature, power, timing, and reliability constraints. Thus DEM can be very complex and challenging to achieve. Since often times many DEMs operate together on a single processor, there is a need to unify various DEM techniques. This dissertation address such a need. In this work, a framework for DEM is proposed that provides a unifying processor model that includes processor power, thermal, timing, and reliability models, supports various DEM control mechanisms, many different objective functions along with equally diverse constraint specifications. Using the framework, a range of novel solutions is derived for instances of DEM problems, that include maximizing processor performance, energy efficiency, or minimizing power consumption, peak temperature under constraints of maximum temperature, memory reliability and task deadlines. Finally, a robust closed-loop controller to implement the above solutions on a real processor platform with a very low operational overhead is proposed. Along with the controller design, a model identification methodology for obtaining the required power and thermal models for the controller is also discussed. The controller is architecture independent and hence easily portable across many platforms. The controller has been successfully deployed on Intel Sandy Bridge processor and the use of the controller has increased the energy efficiency of the processor by over 30%
ContributorsHanumaiah, Vinay (Author) / Vrudhula, Sarma (Thesis advisor) / Chatha, Karamvir (Committee member) / Chakrabarti, Chaitali (Committee member) / Rodriguez, Armando (Committee member) / Askin, Ronald (Committee member) / Arizona State University (Publisher)
Created2013
Description
The advent of threshold logic simplifies the traditional Boolean logic to the single level multi-input function. Threshold logic latch (TLL), among implementations of threshold logic, is functionally equivalent to a multi-input function with an edge triggered flip-flop, which stands out to improve area and both dynamic and leakage power consumption, providing an appropriate design alternative. Accordingly, the TLL standard cell library is designed. Through technology mapping, hybrid circuit is generated by absorbing the logic cone backward from each flip-flip to get the smallest remaining feeder. With the scan test methodology adopted, design for testability (DFT) is proposed, including scan element design and scan chain insertion. Test synthesis flow is then introduced, according to the Cadence tool, RTL compiler. Test application is the process of applying vectors and the response analysis, which is mainly about the testbench design. A parameterized generic self-checking Verilog testbench is designed for static fault detection. Test development refers to the fault modeling, and test generation. Firstly, functional truth table test generation on TLL cells is proposed. Before the truth table test of the threshold function, the dependence of sequence of vectors applied, i.e., the dependence of current state on the previous state, should be eliminated. Transition test (dynamic pattern) on all weak inputs is proved to be able to test the reset function, which is supposed to erase the history in the reset phase before every evaluation phase. Remaining vectors in the truth table except the weak inputs are then applied statically (static pattern). Secondly, dynamic patterns for all weak inputs are proposed to detect structural transistor level faults analyzed in the TLL cell, with single fault assumption and stuck-at faults, stuck-on faults, and stuck-open faults under consideration. Containing those patterns, the functional test covers all testable structural faults inside the TLL. Thirdly, with the scope of the whole hybrid netlist, the procedure of test generation is proposed with three steps: scan chain test; test of feeders and other scan elements except TLLs; functional pattern test of TLL cells. Implementation of this procedure is discussed in the automatic test pattern generation (ATPG) chapter.
ContributorsHu, Yang (Author) / Vrudhula, Sarma (Thesis advisor) / Barnaby, Hugh (Committee member) / Yu, Shimeng (Committee member) / Arizona State University (Publisher)
Created2013
Description
The geometric growth in the integrated circuit technology due to transistor scaling also with system-on-chip design strategy, the complexity of the integrated circuit has increased manifold. Short time to market with high reliability and performance is one of the most competitive challenges. Both custom and ASIC design methodologies have evolved over the time to cope with this but the high manual labor in custom and statistic design in ASIC are still causes of concern. This work proposes a new circuit design strategy that focuses mostly on arrayed structures like TLB, RF, Cache, IPCAM etc. that reduces the manual effort to a great extent and also makes the design regular, repetitive still achieving high performance. The method proposes making the complete design custom schematic but using the standard cells. This requires adding some custom cells to the already exhaustive library to optimize the design for performance. Once schematic is finalized, the designer places these standard cells in a spreadsheet, placing closely the cells in the critical paths. A Perl script then generates Cadence Encounter compatible placement file. The design is then routed in Encounter. Since designer is the best judge of the circuit architecture, placement by the designer will allow achieve most optimal design. Several designs like IPCAM, issue logic, TLB, RF and Cache designs were carried out and the performance were compared against the fully custom and ASIC flow. The TLB, RF and Cache were the part of the HEMES microprocessor.
ContributorsMaurya, Satendra Kumar (Author) / Clark, Lawrence T (Thesis advisor) / Holbert, Keith E. (Committee member) / Vrudhula, Sarma (Committee member) / Allee, David (Committee member) / Arizona State University (Publisher)
Created2012
Description
Residue number systems have gained significant importance in the field of high-speed digital signal processing due to their carry-free nature and speed-up provided by parallelism. The critical aspect in the application of RNS is the selection of the moduli set and the design of the conversion units. There have been several RNS moduli sets proposed for the implementation of digital filters. However, some are unbalanced and some do not provide the required dynamic range. This thesis addresses the drawbacks of existing RNS moduli sets and proposes a new moduli set for efficient implementation of FIR filters. An efficient VLSI implementation model has been derived for the design of a reverse converter from RNS to the conventional two's complement representation. This model facilitates the realization of a reverse converter for better performance with less hardware complexity when compared with the reverse converter designs of the existing balanced 4-moduli sets. Experimental results comparing multiply and accumulate units using RNS that are implemented using the proposed four-moduli set with the state-of-the-art balanced four-moduli sets, show large improvements in area (46%) and power (43%) reduction for various dynamic ranges. RNS FIR filters using the proposed moduli-set and existing balanced 4-moduli set are implemented in RTL and compared for chip area and power and observed 20% improvements. This thesis also presents threshold logic implementation of the reverse converter.
ContributorsChalivendra, Gayathri (Author) / Vrudhula, Sarma (Thesis advisor) / Shrivastava, Aviral (Committee member) / Bakkaloglu, Bertan (Committee member) / Arizona State University (Publisher)
Created2011
Description
This work focuses on the existence of multiple resistance states in a type of emerging non-volatile resistive memory device known commonly as Programmable Metallization Cell (PMC) or Conductive Bridge Random Access Memory (CBRAM), which can be important for applications such as multi-bit memory as well as non-volatile logic and neuromorphic computing. First, experimental data from small signal, quasi-static and pulsed mode electrical characterization of such devices are presented which clearly demonstrate the inherent multi-level resistance programmability property in CBRAM devices. A physics based analytical CBRAM compact model is then presented which simulates the ion-transport dynamics and filamentary growth mechanism that causes resistance change in such devices. Simulation results from the model are fitted to experimental dynamic resistance switching characteristics. The model designed using Verilog-a language is computation-efficient and can be integrated with industry standard circuit simulation tools for design and analysis of hybrid circuits involving both CMOS and CBRAM devices. Three main circuit applications for CBRAM devices are explored in this work. Firstly, the susceptibility of CBRAM memory arrays to single event induced upsets is analyzed via compact model simulation and experimental heavy ion testing data that show possibility of both high resistance to low resistance and low resistance to high resistance transitions due to ion strikes. Next, a non-volatile sense amplifier based flip-flop architecture is proposed which can help make leakage power consumption negligible by allowing complete shutdown of power supply while retaining its output data in CBRAM devices. Reliability and energy consumption of the flip-flop circuit for different CBRAM low resistance levels and supply voltage values are analyzed and compared to CMOS designs. Possible extension of this architecture for threshold logic function computation using the CBRAM devices as re-configurable resistive weights is also discussed. Lastly, Spike timing dependent plasticity (STDP) based gradual resistance change behavior in CBRAM device fabricated in back-end-of-line on a CMOS die containing integrate and fire CMOS neuron circuits is demonstrated for the first time which indicates the feasibility of using CBRAM devices as electronic synapses in spiking neural network hardware implementations for non-Boolean neuromorphic computing.
ContributorsMahalanabis, Debayan (Author) / Barnaby, Hugh J. (Thesis advisor) / Kozicki, Michael N. (Committee member) / Vrudhula, Sarma (Committee member) / Yu, Shimeng (Committee member) / Arizona State University (Publisher)
Created2015
Description
Static CMOS logic has remained the dominant design style of digital systems for
more than four decades due to its robustness and near zero standby current. Static
CMOS logic circuits consist of a network of combinational logic cells and clocked sequential
elements, such as latches and flip-flops that are used for sequencing computations
over time. The majority of the digital design techniques to reduce power, area, and
leakage over the past four decades have focused almost entirely on optimizing the
combinational logic. This work explores alternate architectures for the flip-flops for
improving the overall circuit performance, power and area. It consists of three main
sections.
First, is the design of a multi-input configurable flip-flop structure with embedded
logic. A conventional D-type flip-flop may be viewed as realizing an identity function,
in which the output is simply the value of the input sampled at the clock edge. In
contrast, the proposed multi-input flip-flop, named PNAND, can be configured to
realize one of a family of Boolean functions called threshold functions. In essence,
the PNAND is a circuit implementation of the well-known binary perceptron. Unlike
other reconfigurable circuits, a PNAND can be configured by simply changing the
assignment of signals to its inputs. Using a standard cell library of such gates, a technology
mapping algorithm can be applied to transform a given netlist into one with
an optimal mixture of conventional logic gates and threshold gates. This approach
was used to fabricate a 32-bit Wallace Tree multiplier and a 32-bit booth multiplier
in 65nm LP technology. Simulation and chip measurements show more than 30%
improvement in dynamic power and more than 20% reduction in core area.
The functional yield of the PNAND reduces with geometry and voltage scaling.
The second part of this research investigates the use of two mechanisms to improve
the robustness of the PNAND circuit architecture. One is the use of forward and reverse body biases to change the device threshold and the other is the use of RRAM
devices for low voltage operation.
The third part of this research focused on the design of flip-flops with non-volatile
storage. Spin-transfer torque magnetic tunnel junctions (STT-MTJ) are integrated
with both conventional D-flipflop and the PNAND circuits to implement non-volatile
logic (NVL). These non-volatile storage enhanced flip-flops are able to save the state of
system locally when a power interruption occurs. However, manufacturing variations
in the STT-MTJs and in the CMOS transistors significantly reduce the yield, leading
to an overly pessimistic design and consequently, higher energy consumption. A
detailed analysis of the design trade-offs in the driver circuitry for performing backup
and restore, and a novel method to design the energy optimal driver for a given yield is
presented. Efficient designs of two nonvolatile flip-flop (NVFF) circuits are presented,
in which the backup time is determined on a per-chip basis, resulting in minimizing
the energy wastage and satisfying the yield constraint. To achieve a yield of 98%,
the conventional approach would have to expend nearly 5X more energy than the
minimum required, whereas the proposed tunable approach expends only 26% more
energy than the minimum. A non-volatile threshold gate architecture NV-TLFF are
designed with the same backup and restore circuitry in 65nm technology. The embedded
logic in NV-TLFF compensates performance overhead of NVL. This leads to the
possibility of zero-overhead non-volatile datapath circuits. An 8-bit multiply-and-
accumulate (MAC) unit is designed to demonstrate the performance benefits of the
proposed architecture. Based on the results of HSPICE simulations, the MAC circuit
with the proposed NV-TLFF cells is shown to consume at least 20% less power and
area as compared to the circuit designed with conventional DFFs, without sacrificing
any performance.
more than four decades due to its robustness and near zero standby current. Static
CMOS logic circuits consist of a network of combinational logic cells and clocked sequential
elements, such as latches and flip-flops that are used for sequencing computations
over time. The majority of the digital design techniques to reduce power, area, and
leakage over the past four decades have focused almost entirely on optimizing the
combinational logic. This work explores alternate architectures for the flip-flops for
improving the overall circuit performance, power and area. It consists of three main
sections.
First, is the design of a multi-input configurable flip-flop structure with embedded
logic. A conventional D-type flip-flop may be viewed as realizing an identity function,
in which the output is simply the value of the input sampled at the clock edge. In
contrast, the proposed multi-input flip-flop, named PNAND, can be configured to
realize one of a family of Boolean functions called threshold functions. In essence,
the PNAND is a circuit implementation of the well-known binary perceptron. Unlike
other reconfigurable circuits, a PNAND can be configured by simply changing the
assignment of signals to its inputs. Using a standard cell library of such gates, a technology
mapping algorithm can be applied to transform a given netlist into one with
an optimal mixture of conventional logic gates and threshold gates. This approach
was used to fabricate a 32-bit Wallace Tree multiplier and a 32-bit booth multiplier
in 65nm LP technology. Simulation and chip measurements show more than 30%
improvement in dynamic power and more than 20% reduction in core area.
The functional yield of the PNAND reduces with geometry and voltage scaling.
The second part of this research investigates the use of two mechanisms to improve
the robustness of the PNAND circuit architecture. One is the use of forward and reverse body biases to change the device threshold and the other is the use of RRAM
devices for low voltage operation.
The third part of this research focused on the design of flip-flops with non-volatile
storage. Spin-transfer torque magnetic tunnel junctions (STT-MTJ) are integrated
with both conventional D-flipflop and the PNAND circuits to implement non-volatile
logic (NVL). These non-volatile storage enhanced flip-flops are able to save the state of
system locally when a power interruption occurs. However, manufacturing variations
in the STT-MTJs and in the CMOS transistors significantly reduce the yield, leading
to an overly pessimistic design and consequently, higher energy consumption. A
detailed analysis of the design trade-offs in the driver circuitry for performing backup
and restore, and a novel method to design the energy optimal driver for a given yield is
presented. Efficient designs of two nonvolatile flip-flop (NVFF) circuits are presented,
in which the backup time is determined on a per-chip basis, resulting in minimizing
the energy wastage and satisfying the yield constraint. To achieve a yield of 98%,
the conventional approach would have to expend nearly 5X more energy than the
minimum required, whereas the proposed tunable approach expends only 26% more
energy than the minimum. A non-volatile threshold gate architecture NV-TLFF are
designed with the same backup and restore circuitry in 65nm technology. The embedded
logic in NV-TLFF compensates performance overhead of NVL. This leads to the
possibility of zero-overhead non-volatile datapath circuits. An 8-bit multiply-and-
accumulate (MAC) unit is designed to demonstrate the performance benefits of the
proposed architecture. Based on the results of HSPICE simulations, the MAC circuit
with the proposed NV-TLFF cells is shown to consume at least 20% less power and
area as compared to the circuit designed with conventional DFFs, without sacrificing
any performance.
ContributorsYang, Jinghua (Author) / Vrudhula, Sarma (Thesis advisor) / Barnaby, Hugh (Committee member) / Cao, Yu (Committee member) / Seo, Jae-Sun (Committee member) / Arizona State University (Publisher)
Created2018
Description
The rapid improvement in computation capability has made deep convolutional neural networks (CNNs) a great success in recent years on many computer vision tasks with significantly improved accuracy. During the inference phase, many applications demand low latency processing of one image with strict power consumption requirement, which reduces the efficiency of GPU and other general-purpose platform, bringing opportunities for specific acceleration hardware, e.g. FPGA, by customizing the digital circuit specific for the deep learning algorithm inference. However, deploying CNNs on portable and embedded systems is still challenging due to large data volume, intensive computation, varying algorithm structures, and frequent memory accesses. This dissertation proposes a complete design methodology and framework to accelerate the inference process of various CNN algorithms on FPGA hardware with high performance, efficiency and flexibility.
As convolution contributes most operations in CNNs, the convolution acceleration scheme significantly affects the efficiency and performance of a hardware CNN accelerator. Convolution involves multiply and accumulate (MAC) operations with four levels of loops. Without fully studying the convolution loop optimization before the hardware design phase, the resulting accelerator can hardly exploit the data reuse and manage data movement efficiently. This work overcomes these barriers by quantitatively analyzing and optimizing the design objectives (e.g. memory access) of the CNN accelerator based on multiple design variables. An efficient dataflow and hardware architecture of CNN acceleration are proposed to minimize the data communication while maximizing the resource utilization to achieve high performance.
Although great performance and efficiency can be achieved by customizing the FPGA hardware for each CNN model, significant efforts and expertise are required leading to long development time, which makes it difficult to catch up with the rapid development of CNN algorithms. In this work, we present an RTL-level CNN compiler that automatically generates customized FPGA hardware for the inference tasks of various CNNs, in order to enable high-level fast prototyping of CNNs from software to FPGA and still keep the benefits of low-level hardware optimization. First, a general-purpose library of RTL modules is developed to model different operations at each layer. The integration and dataflow of physical modules are predefined in the top-level system template and reconfigured during compilation for a given CNN algorithm. The runtime control of layer-by-layer sequential computation is managed by the proposed execution schedule so that even highly irregular and complex network topology, e.g. GoogLeNet and ResNet, can be compiled. The proposed methodology is demonstrated with various CNN algorithms, e.g. NiN, VGG, GoogLeNet and ResNet, on two different standalone FPGAs achieving state-of-the art performance.
Based on the optimized acceleration strategy, there are still a lot of design options, e.g. the degree and dimension of computation parallelism, the size of on-chip buffers, and the external memory bandwidth, which impact the utilization of computation resources and data communication efficiency, and finally affect the performance and energy consumption of the accelerator. The large design space of the accelerator makes it impractical to explore the optimal design choice during the real implementation phase. Therefore, a performance model is proposed in this work to quantitatively estimate the accelerator performance and resource utilization. By this means, the performance bottleneck and design bound can be identified and the optimal design option can be explored early in the design phase.
As convolution contributes most operations in CNNs, the convolution acceleration scheme significantly affects the efficiency and performance of a hardware CNN accelerator. Convolution involves multiply and accumulate (MAC) operations with four levels of loops. Without fully studying the convolution loop optimization before the hardware design phase, the resulting accelerator can hardly exploit the data reuse and manage data movement efficiently. This work overcomes these barriers by quantitatively analyzing and optimizing the design objectives (e.g. memory access) of the CNN accelerator based on multiple design variables. An efficient dataflow and hardware architecture of CNN acceleration are proposed to minimize the data communication while maximizing the resource utilization to achieve high performance.
Although great performance and efficiency can be achieved by customizing the FPGA hardware for each CNN model, significant efforts and expertise are required leading to long development time, which makes it difficult to catch up with the rapid development of CNN algorithms. In this work, we present an RTL-level CNN compiler that automatically generates customized FPGA hardware for the inference tasks of various CNNs, in order to enable high-level fast prototyping of CNNs from software to FPGA and still keep the benefits of low-level hardware optimization. First, a general-purpose library of RTL modules is developed to model different operations at each layer. The integration and dataflow of physical modules are predefined in the top-level system template and reconfigured during compilation for a given CNN algorithm. The runtime control of layer-by-layer sequential computation is managed by the proposed execution schedule so that even highly irregular and complex network topology, e.g. GoogLeNet and ResNet, can be compiled. The proposed methodology is demonstrated with various CNN algorithms, e.g. NiN, VGG, GoogLeNet and ResNet, on two different standalone FPGAs achieving state-of-the art performance.
Based on the optimized acceleration strategy, there are still a lot of design options, e.g. the degree and dimension of computation parallelism, the size of on-chip buffers, and the external memory bandwidth, which impact the utilization of computation resources and data communication efficiency, and finally affect the performance and energy consumption of the accelerator. The large design space of the accelerator makes it impractical to explore the optimal design choice during the real implementation phase. Therefore, a performance model is proposed in this work to quantitatively estimate the accelerator performance and resource utilization. By this means, the performance bottleneck and design bound can be identified and the optimal design option can be explored early in the design phase.
ContributorsMa, Yufei (Author) / Vrudhula, Sarma (Thesis advisor) / Seo, Jae-Sun (Thesis advisor) / Cao, Yu (Committee member) / Barnaby, Hugh (Committee member) / Arizona State University (Publisher)
Created2018
Description
Deep learning (DL) has proved itself be one of the most important developements till date with far reaching impacts in numerous fields like robotics, computer vision, surveillance, speech processing, machine translation, finance, etc. They are now widely used for countless applications because of their ability to generalize real world data, robustness to noise in previously unseen data and high inference accuracy. With the ability to learn useful features from raw sensor data, deep learning algorithms have out-performed tradinal AI algorithms and pushed the boundaries of what can be achieved with AI. In this work, we demonstrate the power of deep learning by developing a neural network to automatically detect cough instances from audio recorded in un-constrained environments. For this, 24 hours long recordings from 9 dierent patients is collected and carefully labeled by medical personel. A pre-processing algorithm is proposed to convert event based cough dataset to a more informative dataset with start and end of coughs and also introduce data augmentation for regularizing the training procedure. The proposed neural network achieves 92.3% leave-one-out accuracy on data captured in real world.
Deep neural networks are composed of multiple layers that are compute/memory intensive. This makes it difficult to execute these algorithms real-time with low power consumption using existing general purpose computers. In this work, we propose hardware accelerators for a traditional AI algorithm based on random forest trees and two representative deep convolutional neural networks (AlexNet and VGG). With the proposed acceleration techniques, ~ 30x performance improvement was achieved compared to CPU for random forest trees. For deep CNNS, we demonstrate that much higher performance can be achieved with architecture space exploration using any optimization algorithms with system level performance and area models for hardware primitives as inputs and goal of minimizing latency with given resource constraints. With this method, ~30GOPs performance was achieved for Stratix V FPGA boards.
Hardware acceleration of DL algorithms alone is not always the most ecient way and sucient to achieve desired performance. There is a huge headroom available for performance improvement provided the algorithms are designed keeping in mind the hardware limitations and bottlenecks. This work achieves hardware-software co-optimization for Non-Maximal Suppression (NMS) algorithm. Using the proposed algorithmic changes and hardware architecture
With CMOS scaling coming to an end and increasing memory bandwidth bottlenecks, CMOS based system might not scale enough to accommodate requirements of more complicated and deeper neural networks in future. In this work, we explore RRAM crossbars and arrays as compact, high performing and energy efficient alternative to CMOS accelerators for deep learning training and inference. We propose and implement RRAM periphery read and write circuits and achieved ~3000x performance improvement in online dictionary learning compared to CPU.
This work also examines the realistic RRAM devices and their non-idealities. We do an in-depth study of the effects of RRAM non-idealities on inference accuracy when a pretrained model is mapped to RRAM based accelerators. To mitigate this issue, we propose Random Sparse Adaptation (RSA), a novel scheme aimed at tuning the model to take care of the faults of the RRAM array on which it is mapped. Our proposed method can achieve inference accuracy much higher than what traditional Read-Verify-Write (R-V-W) method could achieve. RSA can also recover lost inference accuracy 100x ~ 1000x faster compared to R-V-W. Using 32-bit high precision RSA cells, we achieved ~10% higher accuracy using fautly RRAM arrays compared to what can be achieved by mapping a deep network to an 32 level RRAM array with no variations.
Deep neural networks are composed of multiple layers that are compute/memory intensive. This makes it difficult to execute these algorithms real-time with low power consumption using existing general purpose computers. In this work, we propose hardware accelerators for a traditional AI algorithm based on random forest trees and two representative deep convolutional neural networks (AlexNet and VGG). With the proposed acceleration techniques, ~ 30x performance improvement was achieved compared to CPU for random forest trees. For deep CNNS, we demonstrate that much higher performance can be achieved with architecture space exploration using any optimization algorithms with system level performance and area models for hardware primitives as inputs and goal of minimizing latency with given resource constraints. With this method, ~30GOPs performance was achieved for Stratix V FPGA boards.
Hardware acceleration of DL algorithms alone is not always the most ecient way and sucient to achieve desired performance. There is a huge headroom available for performance improvement provided the algorithms are designed keeping in mind the hardware limitations and bottlenecks. This work achieves hardware-software co-optimization for Non-Maximal Suppression (NMS) algorithm. Using the proposed algorithmic changes and hardware architecture
With CMOS scaling coming to an end and increasing memory bandwidth bottlenecks, CMOS based system might not scale enough to accommodate requirements of more complicated and deeper neural networks in future. In this work, we explore RRAM crossbars and arrays as compact, high performing and energy efficient alternative to CMOS accelerators for deep learning training and inference. We propose and implement RRAM periphery read and write circuits and achieved ~3000x performance improvement in online dictionary learning compared to CPU.
This work also examines the realistic RRAM devices and their non-idealities. We do an in-depth study of the effects of RRAM non-idealities on inference accuracy when a pretrained model is mapped to RRAM based accelerators. To mitigate this issue, we propose Random Sparse Adaptation (RSA), a novel scheme aimed at tuning the model to take care of the faults of the RRAM array on which it is mapped. Our proposed method can achieve inference accuracy much higher than what traditional Read-Verify-Write (R-V-W) method could achieve. RSA can also recover lost inference accuracy 100x ~ 1000x faster compared to R-V-W. Using 32-bit high precision RSA cells, we achieved ~10% higher accuracy using fautly RRAM arrays compared to what can be achieved by mapping a deep network to an 32 level RRAM array with no variations.
ContributorsMohanty, Abinash (Author) / Cao, Yu (Thesis advisor) / Seo, Jae-Sun (Committee member) / Vrudhula, Sarma (Committee member) / Chakrabarti, Chaitali (Committee member) / Arizona State University (Publisher)
Created2018