Designing energy-efficient computing systems using equalization and machine learning
MetadataShow full item record
As technology scaling slows down in the nanometer CMOS regime and mobile computing becomes more ubiquitous, designing energy-efficient hardware for mobile systems is becoming increasingly critical and challenging. Although various approaches like near-threshold computing (NTC), aggressive voltage scaling with shadow latches, etc. have been proposed to get the most out of limited battery life, there is still no “silver bullet” to increasing power-performance demands of the mobile systems. Moreover, given that a mobile system could operate in a variety of environmental conditions, like different temperatures, have varying performance requirements, etc., there is a growing need for designing tunable/reconfigurable systems in order to achieve energy-efficient operation. In this work we propose to address the energy- efficiency problem of mobile systems using two different approaches: circuit tunability and distributed adaptive algorithms. Inspired by the communication systems, we developed feedback equalization based digital logic that changes the threshold of its gates based on the input pattern. We showed that feedback equalization in static complementary CMOS logic enabled up to 20% reduction in energy dissipation while maintaining the performance metrics. We also achieved 30% reduction in energy dissipation for pass-transistor digital logic (PTL) with equalization while maintaining performance. In addition, we proposed a mechanism that leverages feedback equalization techniques to achieve near optimal operation of static complementary CMOS logic blocks over the entire voltage range from near threshold supply voltage to nominal supply voltage. Using energy-delay product (EDP) as a metric we analyzed the use of the feedback equalizer as part of various sequential computational blocks. Our analysis shows that for near-threshold voltage operation, when equalization was used, we can improve the operating frequency by up to 30%, while the energy increase was less than 15%, with an overall EDP reduction of ≈10%. We also observe an EDP reduction of close to 5% across entire above-threshold voltage range. On the distributed adaptive algorithm front, we explored energy-efficient hardware implementation of machine learning algorithms. We proposed an adaptive classifier that leverages the wide variability in data complexity to enable energy-efficient data classification operations for mobile systems. Our approach takes advantage of varying classification hardness across data to dynamically allocate resources and improve energy efficiency. On average, our adaptive classifier is ≈100× more energy efficient but has ≈1% higher error rate than a complex radial basis function classifier and is ≈10× less energy efficient but has ≈40% lower error rate than a simple linear classifier across a wide range of classification data sets. We also developed a field of groves (FoG) implementation of random forests (RF) that achieves an accuracy comparable to Convolutional Neural Networks (CNN) and Support Vector Machines (SVM) under tight energy budgets. The FoG architecture takes advantage of the fact that in random forests a small portion of the weak classifiers (decision trees) might be sufficient to achieve high statistical performance. By dividing the random forest into smaller forests (Groves), and conditionally executing the rest of the forest, FoG is able to achieve much higher energy efficiency levels for comparable error rates. We also take advantage of the distributed nature of the FoG to achieve high level of parallelism. Our evaluation shows that at maximum achievable accuracies FoG consumes ≈1.48×, ≈24×, ≈2.5×, and ≈34.7× lower energy per classification compared to conventional RF, SVM-RBF , Multi-Layer Perceptron Network (MLP), and CNN, respectively. FoG is 6.5× less energy efficient than SVM-LR, but achieves 18% higher accuracy on average across all considered datasets.