Publication: Evaluación del SoC M1 de Apple para cómputo matricial
Loading...
Official URL
Full text at PDC
Publication Date
2022
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
El cómputo matricial está presente en multitud de ámbitos de la vida cotidiana. Una de las aplicaciones donde está siendo más utilizado en el Machine Learning, donde el tratamiento de volúmenes de datos grandes es frecuente.
En este proyecto se evalúan las ventajas que presentan las distintas librerías BLAS tanto a la hora de crear redes neuronales como a la hora de realizar operaciones algebraicas básicas, que son la base del Machine Learning.
Se han analizado diversos aspectos de estas librerías como el rendimiento, el consumo y la eficiencia de las librerías Accelerate, OpenBLAS y BLIS y se han obtenido conclusiones acerca de su viabilidad para el uso en aplicaciones de Machine Learning. Además, se analiza el rendimiento de la librería Accelerate, que es propia de Apple, con la librería MetalPerfomanceShaders, que es también propiedad de la misma compañía y permite el acceso a la GPU.
Las pruebas se han ejecutado sobre el modelo más actual de Mac Mini, que incluye el procesador Apple M1 de última generación.
Matrix computation is present in several daily life areas. One of the areas where its being more used is Machine Learning, since big-volume data processing is frequent. In this project, several advantages of the BLAS libraries are analized when it comes to creating neural networks and performing basic algebraic operations, which are the base of Machine Learning. Multiple aspects such as performance, energy consumption and efficiency have been analized from the libraries Accelerate, OpenBLAS and BLIS and conclusions have been obtained regarding their viability when it comes to using them in Machine Learning applications. Also, the performance of the Accelerate Apple-propietary library has been analized against the performance obtained with the also-Apple-propietary MetalPerformanceShaders library, which allows the GPU usage. These tests have been performed on the latest Apple Mac Mini, which includes the latest Apple M1 chip.
Matrix computation is present in several daily life areas. One of the areas where its being more used is Machine Learning, since big-volume data processing is frequent. In this project, several advantages of the BLAS libraries are analized when it comes to creating neural networks and performing basic algebraic operations, which are the base of Machine Learning. Multiple aspects such as performance, energy consumption and efficiency have been analized from the libraries Accelerate, OpenBLAS and BLIS and conclusions have been obtained regarding their viability when it comes to using them in Machine Learning applications. Also, the performance of the Accelerate Apple-propietary library has been analized against the performance obtained with the also-Apple-propietary MetalPerformanceShaders library, which allows the GPU usage. These tests have been performed on the latest Apple Mac Mini, which includes the latest Apple M1 chip.
Description
Trabajo de Fin de Grado en Ingeniería de Computadores, Facultad de Informática UCM, Departamento de Arquitectura de Computadores y Automática, Curso 2021/2022.