Work

Approximate Computing for Aging Compensation and Energy-efficient Neural Network

Year: 2020
Type: dissertation
Abstract: Approximate computing reduces the cost (energy and/or latency) of computations by relaxing the correctness (i.e., precision) of computations up to the level, which is dependent on types of applications. Moreover, it can be realized in various hierarchies of computing system design from circuit level to application level. \n This dissertation presents the methodologies applying approximate computing across such hierarchies; compensating aging-induced delay in logic circuit by dynamic computation approximation (Chapter 1), designing energy-efficient neural network by combining low-power and low-latency approximate neuron models (Chapter 2), and co-designing in-memory gradient descent module with neural processing unit so as to address a memory bottleneck incurred by memory I/O for high-precision data (Chapter 3). \n The first chapter of this dissertation presents a novel design methodology to turn the timing violation caused by aging into computation approximation error without the reliability guardband or increasing the supply voltage. It can be realized by accurately monitoring the critical path delay at run-time. The proposal is evaluated at two levels: RTL component level and system level. The experimental results at the RTL component level show a significant improvement in terms of (normalized) mean squared error caused by the timing violation and, at the system level, show that the proposed approach successfully transforms the aging-induced timing violation errors into much less harmful computation approximation errors, therefore it recovers image quality up to perceptually acceptable levels. It reduces the dynamic and static power consumption by 21.45% and 10.78%, respectively, with 0.8% area overhead compared to the conventional approach. \n The second chapter of this dissertation presents an energy-efficient neural network consisting of alternative neuron models; Stochastic-Computing (SC) and Spiking (SP) neuron models. SC has been adopted in various fields to improve the power efficiency of systems by performing arithmetic computations stochastically, which approximates binary computation in conventional computing systems. Moreover, a recent work showed that deep neural network (DNN) can be implemented in the manner of stochastic computing and it greatly reduces power consumption. However, Stochastic DNN (SC-DNN) suffers from problem of high latency as it processes only a bit per cycle. To address such problem, it is proposed to adopt Spiking DNN (SP-DNN) as an input interface for SC-DNN since SP effectively processes more bits per cycle than SC-DNN. Moreover, this chapter resolves the encoding mismatch problem, between two different neuron models, without hardware cost by compensating the encoding mismatch with synapse weight calibration. A resultant hybrid DNN (SPSC-DNN) consists of SP-DNN as bottom layers and SC-DNN as top layers. Exploiting the reduced latency from SP-DNN and low-power consumption from SC-DNN, the proposed SPSC-DNN achieves improved energy-efficiency with lower error-rate compared to SC-DNN and SP-DNN in same network configuration. \n The third chapter of this dissertation proposes GradPim architecture, which accelerates the parameter updates by in-memory processing which is codesigned with 8-bit floating-point training in Neural Processing Unit (NPU) for deep neural networks. By keeping the high precision processing algorithms in memory, such as the parameter update incorporating high-precision weights in its computation, the GradPim architecture can achieve high computational efficiency using 8-bit floating point in NPU and also gain power efficiency by eliminating massive high-precision data transfers between NPU and off-chip memory. A simple extension of DDR4 SDRAM utilizing bank-group parallelism makes the operation designs in processing-in-memory (PIM) module efficient in terms of hardware cost and performance. The experimental results show that the proposed architecture can improve the performance of the parameter update phase in the training by up to 40% and greatly reduce the memory bandwidth requirement while posing only a minimal amount of overhead to the protocol and the DRAM area.
Author: 김희수
Language: English
Open Access status: green
OpenAlex ID: W3111709921
Cites: 113
Cited by:
Related to: 10