Optimización de un acelerador para multiplicación matricial mediante computación aproximada
Loading...
Date
Authors
Cerdas-Mora, Carlos Adrián
Journal Title
Journal ISSN
Volume Title
Publisher
Instituto Tecnológico de Costa Rica
Abstract
La Inteligencia Artificial (IA) ha transformado la tecnología, al mejorar la eficiencia en
diversas industrias. Sin embargo, las aplicaciones de IA requieren un alto rendimiento
computacional, lo que plantea desafíos de recursos y consumo energético. Los aceleradores
de hardware especializados, como los de multiplicación de matrices, mejoran el
rendimiento y la eficiencia energética.
El trabajo se enfoca en investigación y puesta en práctica de técnicas para computación
aproximada, como el truncamiento y la cuantización, con optimización en la multiplicación
de matrices en sistemas de IA. El objetivo es conocer la mejor aproximación que equilibre
eficiencia y precisión de los recursos disponibles para 4 distintos niveles de profundidad
en cada tipo de diseño. En particular, la aproximación conjunta con 4 bits mostró ser la
mejor opción entre todas las configuraciones evaluadas. Este diseño maximiza la eficiencia
computacional y mantiene un nivel de error aceptable, lo que la convierte en una opción
ideal para diversas aplicaciones que requieren un buen equilibrio entre rendimiento y
precisión, como en sistemas embebidos o dispositivos de bajo consumo.
Artificial Intelligence (AI) has transformed technology by improving efficiency in various industries. However, AI applications require high computational performance, which poses resource and power consumption challenges. Specialized hardware accelerators, such as matrix multiplication accelerators, improve performance and energy efficiency. The work focuses on investigating and evaluating approximate computing techniques, such as truncation and quantization, to optimize matrix multiplication in AI systems. The objective is to find the best approximation that balances efficiency and accuracy as a function of available resources for 4 different depth levels in each type of approximation. In particular, the joint approximation with 4 bits was shown to be the best choice among all the evaluated configurations. This combination maximizes computational efficiency and maintains an acceptable error level, making it an ideal choice for various applications that require a good balance between performance and accuracy, such as in embedded systems or low-power devices.
Artificial Intelligence (AI) has transformed technology by improving efficiency in various industries. However, AI applications require high computational performance, which poses resource and power consumption challenges. Specialized hardware accelerators, such as matrix multiplication accelerators, improve performance and energy efficiency. The work focuses on investigating and evaluating approximate computing techniques, such as truncation and quantization, to optimize matrix multiplication in AI systems. The objective is to find the best approximation that balances efficiency and accuracy as a function of available resources for 4 different depth levels in each type of approximation. In particular, the joint approximation with 4 bits was shown to be the best choice among all the evaluated configurations. This combination maximizes computational efficiency and maintains an acceptable error level, making it an ideal choice for various applications that require a good balance between performance and accuracy, such as in embedded systems or low-power devices.
Description
Proyecto de Graduación (Licenciatura en Ingeniería Electrónica) Instituto Tecnológico de Costa Rica, Escuela de Ingeniería Electrónica, 2024.
Keywords
Optimización -- Aceleradores, Computación, Aceleradores de multiplicación de matrices, Inteligencia artificial, Eficiencia energética, Sistemas empotrados, Arquitectura de computadores, Desarrollo tecnológico, Optimization -- Accelerators, Computing, Matrix multiplication accelerators, Artificial intelligence, Energy efficiency, Embedded systems, Computer architecture, Technological development, Research Subject Categories::TECHNOLOGY::Electrical engineering, electronics and photonics, Research Subject Categories::TECHNOLOGY::Information technology::Computer science
Citation
Collections
Endorsement
Review
Supplemented By
Referenced By
Creative Commons license
Except where otherwised noted, this item's license is described as acceso abierto
