TBA Co-sponsored by: Wright-Patt Multi-Intelligence Development Consortium (WPMDC), The DOD & DOE Communities Speaker(s): Mark Bateman Agenda: TBA Virtual: https://events.vtools.ieee.org/m/422290
Training Deep Neural Networks (DNNs) requires a large number of operations, among which matrix-vector multiplies (MVMs), often of high dimensionality, dominate. In-Memory Computing (IMC) is a promising approach to enhance MVM compute efficiency and throughput, but introduces fundamental tradeoffs with dynamic range of the computed outputs. While IMC has been successful in DNN inference systems, it has not yet shown feasibility for training, which is more sensitive to dynamic range. This work leverages recent work on alternative radix-4 number formats in DNN training on digital architectures, together with recent work on high-precision analog IMC with multi-level inputs, to enable IMC training. Furthermore, we implement a mapping of radix-4 operands to multi-level analog-input IMC in a manner that improves robustness to analog noise effects. The proposed approach is shown in simulations calibrated to silicon-measured IMC noise to be capable of training DNNs on the CIFAR-10 dataset to within 10% of the testing accuracy of standard DNN training approaches, while analysis shows that further reduction of IMC noise to feasible levels results in accuracy within 2% of standard DNN training approaches. Co-sponsored by: Wright-Patt Multi-Intelligence Development Consortium (WPMDC), The DOD & DOE Communities Speaker(s): Christopher Grimm Agenda: Training Deep Neural Networks (DNNs) requires a large number of operations, among which matrix-vector multiplies (MVMs), often of high dimensionality, dominate. In-Memory Computing (IMC) is a promising approach to enhance MVM compute efficiency and throughput, but introduces fundamental tradeoffs with dynamic range of the computed outputs. While IMC has been successful in DNN inference systems, it has not yet shown feasibility for training, which is more sensitive to dynamic range. This work leverages recent work on alternative radix-4 number formats in DNN training on digital architectures, together with recent work on high-precision analog IMC with multi-level inputs, to enable IMC training. Furthermore, we implement a mapping of radix-4 operands to multi-level analog-input IMC in a manner that improves robustness to analog noise effects. The proposed approach is shown in simulations calibrated to silicon-measured IMC noise to be capable of training DNNs on the CIFAR-10 dataset to within 10% of the testing accuracy of standard DNN training approaches, while analysis shows that further reduction of IMC noise to feasible levels results in accuracy within 2% of standard DNN training approaches. Virtual: https://events.vtools.ieee.org/m/422920
Please join us for the IEEE Dayton Section Quarterly Meeting & Society/Affinity Group Chair Updates on Thursday June 20, 2024! The meeting will be held at the Wandering Griffin restaurant from 11:30 am to 1:00 pm. The address is: 3725 Presential Dr., Beavercreek, OH 45324 (https://www.wanderinggriffin.com/) LUNCH WILL BE PROVIDED! (up to $15) NOTE: Society and Affinity Group Chairs will be asked to provide an update on year-to-date activities. If you cannot attend the meeting in person, a remote connection option will be provided (details coming soon). You can also email your information (along with any questions) in advance to: Amy Doll, Vice Chair, IEEE Dayton Section at [email protected] Wandering Griffin Brewery, 3725 Presidential Drive, Beavercreek, Ohio, United States, 45324
The demand for faster communication technologies has motived exploration of higher frequency bands. Positioned between the microwave and infrared regions of the electromagnetic spectrum, the THz band offers the potential for ultra-high data rates and significant bandwidth. However, its practical utilization faces substantial challenges, including a limited number of available sources and receivers and severe atmospheric attenuation, collectively known as the "THz Gap." This research aims to address these challenges by exploring THz generation concepts and determining the feasibility of real-time propagation predictions. The study primarily focuses on using superconducting ring arrays to generate Radio Frequency (RF) radiation, leveraging prior and ongoing research. Characterization of THz emission will be conducted through modeling and measurements, primarily using FEKO, a commercially available electromagnetic solver. Measurements will utilize scaling principles of antenna structures, with efforts to detect actual THz radiation through various approaches. In addition to THz generation, the study will model THz propagation using the High Energy Laser End to End Operations Simulation (HELEEOS) tool developed by the Center for Directed Energy at the Air Force Institute of Technology (AFIT). This MATLAB-implemented code, originally developed for laser propagation predictions, will be applied to the THz frequency range for the first time. The proposed dissertation aims to answer two key research questions: (1) Can the generation of THz radiation using superconducting rings provide a viable alternative to existing approaches? (2) Can RF scaling principles enable the construction and prediction of THz propagation to assess true THz band utilization? The qualification of these questions will be achieved through a combination of measurements, modeling, and quantitative analysis, providing a comprehensive understanding of THz generation and propagation. Co-sponsored by: Wright-Patt Multi-Intelligence Development Consortium (WPMDC), The DOD & DOE Communities Speaker(s): Nathan Agenda: The demand for faster communication technologies has motived exploration of higher frequency bands. Positioned between the microwave and infrared regions of the electromagnetic spectrum, the THz band offers the potential for ultra-high data rates and significant bandwidth. However, its practical utilization faces substantial challenges, including a limited number of available sources and receivers and severe atmospheric attenuation, collectively known as the "THz Gap." This research aims to address these challenges by exploring THz generation concepts and determining the feasibility of real-time propagation predictions. The study primarily focuses on using superconducting ring arrays to generate Radio Frequency (RF) radiation, leveraging prior and ongoing research. Characterization of THz emission will be conducted through modeling and measurements, primarily using FEKO, a commercially available electromagnetic solver. Measurements will utilize scaling principles of antenna structures, with efforts to detect actual THz radiation through various approaches. In addition to THz generation, the study will model THz propagation using the High Energy Laser End to End Operations Simulation (HELEEOS) tool developed by the Center for Directed Energy at the Air Force Institute of Technology (AFIT). This MATLAB-implemented code, originally developed for laser propagation predictions, will be applied to the THz frequency range for the first time. The proposed dissertation aims to answer two key research questions: (1) Can the generation of THz radiation using superconducting rings provide a viable alternative to existing approaches? (2) Can RF scaling principles enable the construction and prediction of THz propagation to assess true THz band utilization? The qualification of these questions will be achieved through a combination of measurements, modeling, and quantitative analysis, providing a comprehensive understanding of THz generation and propagation. Virtual: https://events.vtools.ieee.org/m/425412
