18 research outputs found
Recommended from our members
Design for Ultra Low Power Hardware Accelerator Using Approximate and In-Memory Computing
No full textIn recent years, the rapid slowdown of Moore's Law has made it impossible to enhance processor performance at the same rate by merely scaling down transistor size. Simultaneously, the volume of data processed by computing systems has surged, making the memory wall a significant bottleneck for artificial intelligence (AI) and digital signal processing (DSP) accelerators. Data-intensive computing tasks such as AI and DSP have led to a dramatic rise in energy demands for computing, far outpacing the linear growth of global energy production. Consequently, alternative memory-computing paradigms beyond the Von Neumann architecture, along with low-power and hardware-efficient computing approaches enabled by approximation, are gaining popularity, especially for emerging AI and media applications that can tolerate higher error rates. In this thesis, we try to address the power efficiency of the hardware accelerator designs. First, we propose an efficient hybrid parallel Processing-in-Memory (PIM)-based computation for the Discrete Hadamard Transform (DHT). Our method leverages the recursive computation of DHT using memristor-aided logic (MAGIC) gates, where arithmetic operations are performed via simple logic NOR operations. Second, to further reduce power consumption, and improve computing throughput, we develop a novel multi-level reconfigurable approximate logarithmic multiplier design named Multi-ALM. The Multi-ALM is based on a new iterative formulation of approximate logarithmic multiplication, which is mathematically proven and similar to a Taylor series, to balance accuracy and performance/power. This design offers a new way to trade-off accuracy for power/performance in a systematic and progressive manner. Last but not least, we propose a new hybrid temporal computing framework that combines pulse rate and temporal data encoding for ultra-low energy hardware accelerators. Our approach is inspired by recent advances in temporal computing or race logic, where data values are encoded as single delays, resulting in significantly lower energy consumption due to minimized single switching events
Recommended from our members
Towards Addressing Thermal and Reliability Challenges in Nanometer Integrated Circuits
No full textOn-chip power densities continue to increase in modern integrated circuits (IC) due to rapid integration and feature scaling.As a consequence, today's high-performance processors have become more thermally constrained than ever before. Increase in temperature has been shown to exponentially degrade reliability of semiconductor chips and has consequently become one of the leading concerns in the industry today. In this thesis, we present our findings and share our contributions from our research efforts in the areas of pre-silicon IC reliability analysis, post-silicon thermal estimation, and advanced microprocessor cooling. Specifically, the first segment of this manuscript will focus on a novel structure-based approach to accelerating electromigration (EM) wear-out for the purposes of post-silicon qualification and burn-in testing. The proposed approach achieves time-to-failure acceleration comparable to the existing current and temperature based stressing techniques at close to nominal operating conditions. Temperature and reliability go hand-in-hand; hence monitoring and managing the processor's temperature while it is in use is equally important in order to maximize performance while minimizing reliability impacts. Therefore, the second segment of this thesis will present our data-driven post-silicon approach to estimating the spatial temperature distribution across the surface of the die in real time. This approach leverages the latest advancements in recurrent-neural-networks for time-series estimation. The estimated temperatures from the proposed model can then be used to supplement the temperature information sensed from the embedded thermal sensors in order to make better informed thermal and reliability regulation decisions. Lastly, the third segment of this thesis will focus on leveraging the aforementioned real-time temperature estimation technique and the emerging thermo-electric based active cooling technologies to propose an on-demand targeted cooling system for modern high-performance processors. This approach yields the sub-ambient cooling benefits of thermo-electric cooling with lower power overheads
Recent advance in computational prototyping for analysis of high-performance analog/RF ICs
No full textA survey of RLCK reduction and simulation methods by fast truncated balanced realization
No full text
Frataxin deficiency increases cyclooxygenase 2 and prostaglandins in cell and animal models of Friedreich's ataxia
No full text© The Author 2014. Published by Oxford University Press
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.An inherited deficiency of the mitochondrial protein frataxin causes Friedreich's ataxia (FRDA); the mechanism by which this deficiency triggers neuro- and cardio-degeneration is unclear. Microarrays of neural tissue of animal models of the disease showed decreases in antioxidant genes, and increases in inflammatory genes. Cyclooxygenase (COX)-derived oxylipins are important mediators of inflammation. We measured oxylipin levels using tandem mass spectrometry and ELISAs in multiple cell and animal models of FRDA. Mass spectrometry revealed increases in concentrations of prostaglandins, thromboxane B2, 15-HETE and 11-HETE in cerebellar samples of knockin knockout mice. One possible explanation for the elevated oxylipins is that frataxin deficiency results in increased COX activity. While constitutive COX1 was unchanged, inducible COX2 expression was elevated over 1.35-fold (P < 0.05) in two Friedreich's mouse models and Friedreich's lymphocytes. Consistent with higher COX2 expression, its activity was also increased by 58% over controls. COX2 expression is driven by multiple transcription factors, including activator protein 1 and cAMP response element-binding protein, both of which were elevated over 1.52-fold in cerebella. Taken together, the results support the hypothesis that reduced expression of frataxin leads to elevation of COX2-mediated oxylipin synthesis stimulated by increases in transcription factors that respond to increased reactive oxygen species. These findings support a neuroinflammatory mechanism in FRDA, which has both pathomechanistic and therapeutic implications.The study was supported by NIH grants NS077777, EY012245 and AG025532 to G.A.C., and USDA-ARS Intramural Projects 5306-51530-019-00D and 1 U24 DK097154-01 to J.W.N. Funding to pay the Open Access publication charges for this article was provided by the NIH
