Publication: P-IBEX 2.0: acelerador para posits basado en un core Ibex
Loading...
Official URL
Full text at PDC
Publication Date
2021-09-21
Advisors (or tutors)
Barrio García, Alberto Antonio del
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
Hoy en día se están alcanzando velocidades de cálculo y procesamiento enormes, como se puede observar en los ranking de “Top500” y “Green500”, y los teraflops han aumentado considerablemente. Sin embargo, ¿sería correcto el uso exclusivo de la representación en coma flotante? ¿Se podría encontrar un formato más eficiente para algunos ámbitos?
La representación Unum (Universal Numbers) surge como alternativa de representación de números decimales similar al punto flotante y cuya última iteración, Type III Unum, también conocido como formato Posit, ofrece un formato y una aritmética donde se pueden representar y operar datos. Dicha encapsulación aporta una representación más exacta a la que se conseguiría siguiendo la regla IEEE 754 con el mismo número de bits.
Para poder empezar a considerar este formato primero se de plantean proyectos que demuestren su conveniencia, se propone realizar un coprocesador que ejecute operaciones básicas y someterlo a pruebas para comprobar el rendimiento. Para alcanzar este objetivo, se exponen varias propuestas de coprocesadores que permitan realizar operaciones con números Posit y así como su comparación.
Para llevar a cabo este proyecto se usa el procesador Ibex de LowRISC, procesador open source basado en RISC-V, sencillo y asequible para este trabajo. Por otra parte, se utilizará el diseño e implementación de PACoGen para la aritmética posit, centrándose en las unidades segmentadas.
Ya que el objetivo de este trabajo consiste en aportar más datos y pruebas a la comunidad, se seguirá una filosofía open source, liberalizando el código y documentación que se genere en el desarrollo del mismo.
Por último, se practicarán distintas pruebas para las propuestas del proyecto, comprobando así cual es el más óptimo y su posterior análisis.
Nowadays, computers are reaching computing and processing speeds extremely high, as it is shown in the rankings ”Top500” and “Green500”. Also, teraflops have increased, however, would it be correct the use of the representation of floating point? Would it be possible to find a more efficient format for different areas? The representation of Unum (Universal Numbers) starts as an alternative to the representation of decimal numbers. It is similar to floating point and its last iteration, Type III Unum, also known as Posit format. It offers an arithmetic and a format that allows to represent and to operate several data. This encapsulation contributes to a more exact representation than the one following the rule IEEE 754 with the same number of bits. To consider this format, first of all it is mandatory to propose some projects and to check its viability. To make a coprocessor that runs basic operations and that gives some tests to confirm its efficiency. To reach this objective, there are several proposals of different coprocessors that execute operations with Posit numbers and compare them. In order to accomplish this project, the processor which is being used is Ibex, from LowRISC, an open source and simple processor based on RISC-V. Nevertheless, the design and the implementation will be from PACoGen for the posit arithmetic. Being focus on the segmented units, too.
Nowadays, computers are reaching computing and processing speeds extremely high, as it is shown in the rankings ”Top500” and “Green500”. Also, teraflops have increased, however, would it be correct the use of the representation of floating point? Would it be possible to find a more efficient format for different areas? The representation of Unum (Universal Numbers) starts as an alternative to the representation of decimal numbers. It is similar to floating point and its last iteration, Type III Unum, also known as Posit format. It offers an arithmetic and a format that allows to represent and to operate several data. This encapsulation contributes to a more exact representation than the one following the rule IEEE 754 with the same number of bits. To consider this format, first of all it is mandatory to propose some projects and to check its viability. To make a coprocessor that runs basic operations and that gives some tests to confirm its efficiency. To reach this objective, there are several proposals of different coprocessors that execute operations with Posit numbers and compare them. In order to accomplish this project, the processor which is being used is Ibex, from LowRISC, an open source and simple processor based on RISC-V. Nevertheless, the design and the implementation will be from PACoGen for the posit arithmetic. Being focus on the segmented units, too.
Description
Trabajo de Fin de Grado en Ingeniería Informática, Facultad de Informática UCM, Departamento de Arquitectura de Computadores y Automática, Curso 2020-2021