Post-synthesis manipulation of field-coupled nanocomputing circuits for fundamental energy reduction

• Autor(a) principal: Jeferson Figueiredo Chaves
• Data de Publicação: 2020
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: http://hdl.handle.net/1843/38083
• topic: Nanocomputing; Reversible Computing; Field-coupled nanocomputing devices; Computação – Teses; Sistemas digitais – Teses.; Arquitetura de computador – Teses.; Nanocomputação – Teses.; Computação Reversível – Teses

Progress in computer performance is promoted on several fronts. Initiatives such as the implementation of optimal algorithms, parallelization of algorithms and architectures, and, at a lower level, Moore's Law, have guaranteed performance evolution. The latter, in particular, already shows clear signs of exhaustion, imposing on industry and academia the search for alternatives. The reason is heat generation, which has been one of the main obstacles to reducing digital devices' dimensions. Landauer investigated the origin of this problem and demonstrated analytically in 1961 that the cause is the erasure of information in each logic gate. This fundamental energy limit, named Landauer's Principle, was experimentally proven in 2012, indicating that there is indeed a limit on energy scalability for conventional computer systems. In this work, we introduce approaches to reduce those losses by applying post-synthesis methods. First, we propose a novel method to divide Partially Reversible Pipelined circuits into stages, computing the optimal energy vs. throughput tradeoff. Specifically, we develop an algorithm that, given the circuit, it returns the division that yields the minimum possible energy dissipation while considering a minimum throughput restriction. We also developed an optimal algorithm that, given the maximum energy restriction, finds the maximum possible throughput. Therefore, the designer can have the best partially reversible pipelined circuit configuration according to its project requirements. Computational experiments show that our algorithm outperforms the state-of-the-art algorithm from the literature in solution's quantity and quality, being able to dominate about 56% of its best solutions and to generate 11X more solutions, on average. The other strategy is a novel technique that identifies exploitable fan-outs and uses them to reduce energy losses in FCN circuits, thus enabling the design of new types of partially reversible systems. Moreover, we propose an algorithm that creates partially reversible systems while allowing the designer to choose between energy reduction with no effect on the delay or focus solely on energy and accepting a potential delay penalty. Simulation results for state-of-the-art benchmarks indicate an average reduction of the fundamental energy limit by 44% without affecting the delay. If delay is not the main concern, the average reduction reaches even 77%. Finally, we present a unified analysis of both partially reversible systems. These efforts open new perspectives in designing FCN systems where energy efficiency is mandatory.