Comparison of Vitis-AI and FINN for implementing convolutional neural networks on FPGA

Convolutional neural networks (CNNs) are essential for image classification and detection, and their implementation in embedded systems is becoming increasingly attractive due to their compact size and low power consumption. Field-Programmable Gate Arrays (FPGAs) have  emerged as a promising option,...

Descripción completa

Guardado en:
Detalles Bibliográficos
Autores principales: Urbano Pintos, Nicolás, Lacomi, Héctor, Lavorato, Mario
Formato: Artículo publishedVersion
Lenguaje:Español
Publicado: FIUBA 2024
Materias:
FPGA
CNN
FINN
Vitis-AI
Quantization
Cuantización
Acceso en línea:https://elektron.fi.uba.ar/elektron/article/view/200
https://repositoriouba.sisbi.uba.ar/gsdl/cgi-bin/library.cgi?a=d&c=elektron&d=200_oai
Aporte de:
Repositorio Digital de la Universidad de Buenos Aires (UBA) de Universidad de Buenos Aires
id I28-R145-200_oai
record_format dspace
spelling I28-R145-200_oai2026-02-11 Urbano Pintos, Nicolás Lacomi, Héctor Lavorato, Mario 2024-12-15 Convolutional neural networks (CNNs) are essential for image classification and detection, and their implementation in embedded systems is becoming increasingly attractive due to their compact size and low power consumption. Field-Programmable Gate Arrays (FPGAs) have  emerged as a promising option, thanks to their low latency and high energy efficiency. Vitis AI and FINN are two development environments that automate the implementation of CNNs on FPGAs. Vitis AI uses a deep learning processing unit (DPU) and memory accelerators, while FINN is based on a streaming architecture and fine-tunes parallelization. Both environments implement  parameter quantization techniques to reduce memory usage. This work extends previous comparisons by evaluating both environments by implementing four models with different numbers of layers on the Xilinx Kria KV260 FPGA platform.  The complete process from training to evaluation on FPGA, including quantization and hardware implementation, is described in  detail. The results show that FINN provides lower latency, higher throughput, and better energy efficiency than Vitis AI. However, Vitis AI stands out for its simplicity in model training and ease of implementation on FPGA. The main finding of this study is that as the complexity of the models increases (with more layers in the neural networks), the differences in terms of performance and energy efficiency between FINN and Vitis AI are significantly reduced. Las redes neuronales convolucionales (CNN) son esenciales para la clasificación y detección de imágenes, y su implementación en sistemas embebidos resulta cada vez más atractiva debido a su tamaño compacto y bajo consumo energético. Los FPGA (Field-Programmable Gate Arrays) han surgido como una opción prometedora, gracias a su baja latencia y alta eficiencia energética. Vitis AI y FINN son dos entornos de desarrollo que automatizan la implementación de CNN en FPGA. Vitis AI utiliza una unidad de procesamiento de aprendizaje profundo (DPU) y aceleradores de memoria, mientras que FINN se basa en una arquitectura de transmisión de datos (streaming) y ajusta la paralelización. Ambos entornos implementan técnicas de cuantización de parámetros para reducir el uso de memoria. Este trabajo extiende comparaciones previas al evaluar ambos entornos mediante la implementación de cuatro modelos con diferentes cantidades de capas en la plataforma FPGA Kria KV260 de Xilinx. Se describe en detalle el proceso completo, desde el entrenamiento hasta la evaluación en FPGA, incluyendo la cuantización y la implementación en hardware. Los resultados muestran que FINN proporciona menor latencia, mayor rendimiento y mejor eficiencia energética que Vitis AI. No obstante, Vitis AI destaca por su simplicidad en el entrenamiento de modelos y facilidad de implementación en FPGA. El hallazgo principal del estudio es que, al aumentar la complejidad de los modelos con más capas, las diferencias de rendimiento y eficiencia energética entre FINN y Vitis AI se reducen notablemente. application/pdf text/html https://elektron.fi.uba.ar/elektron/article/view/200 10.37537/rev.elektron.8.2.200.2024 spa FIUBA https://elektron.fi.uba.ar/elektron/article/view/200/357 https://elektron.fi.uba.ar/elektron/article/view/200/369 Derechos de autor 2024 Nicolás Urbano Pintos Elektron Journal; Vol. 8 No. 2 (2024); 61-70 Revista Elektron; Vol. 8 Núm. 2 (2024); 61-70 Revista Elektron; v. 8 n. 2 (2024); 61-70 2525-0159 2525-0159 FPGA CNN FINN Vitis-AI Quantization FPGA CNN FINN Vitis-AI Cuantización Comparison of Vitis-AI and FINN for implementing convolutional neural networks on FPGA Comparación de Vitis-AI y FINN para implementar redes neuronales convolucionales en FPGA info:eu-repo/semantics/article info:eu-repo/semantics/publishedVersion https://repositoriouba.sisbi.uba.ar/gsdl/cgi-bin/library.cgi?a=d&c=elektron&d=200_oai
institution Universidad de Buenos Aires
institution_str I-28
repository_str R-145
collection Repositorio Digital de la Universidad de Buenos Aires (UBA)
language Español
orig_language_str_mv spa
topic FPGA
CNN
FINN
Vitis-AI
Quantization
FPGA
CNN
FINN
Vitis-AI
Cuantización
spellingShingle FPGA
CNN
FINN
Vitis-AI
Quantization
FPGA
CNN
FINN
Vitis-AI
Cuantización
Urbano Pintos, Nicolás
Lacomi, Héctor
Lavorato, Mario
Comparison of Vitis-AI and FINN for implementing convolutional neural networks on FPGA
topic_facet FPGA
CNN
FINN
Vitis-AI
Quantization
FPGA
CNN
FINN
Vitis-AI
Cuantización
description Convolutional neural networks (CNNs) are essential for image classification and detection, and their implementation in embedded systems is becoming increasingly attractive due to their compact size and low power consumption. Field-Programmable Gate Arrays (FPGAs) have  emerged as a promising option, thanks to their low latency and high energy efficiency. Vitis AI and FINN are two development environments that automate the implementation of CNNs on FPGAs. Vitis AI uses a deep learning processing unit (DPU) and memory accelerators, while FINN is based on a streaming architecture and fine-tunes parallelization. Both environments implement  parameter quantization techniques to reduce memory usage. This work extends previous comparisons by evaluating both environments by implementing four models with different numbers of layers on the Xilinx Kria KV260 FPGA platform.  The complete process from training to evaluation on FPGA, including quantization and hardware implementation, is described in  detail. The results show that FINN provides lower latency, higher throughput, and better energy efficiency than Vitis AI. However, Vitis AI stands out for its simplicity in model training and ease of implementation on FPGA. The main finding of this study is that as the complexity of the models increases (with more layers in the neural networks), the differences in terms of performance and energy efficiency between FINN and Vitis AI are significantly reduced.
format Artículo
publishedVersion
author Urbano Pintos, Nicolás
Lacomi, Héctor
Lavorato, Mario
author_facet Urbano Pintos, Nicolás
Lacomi, Héctor
Lavorato, Mario
author_sort Urbano Pintos, Nicolás
title Comparison of Vitis-AI and FINN for implementing convolutional neural networks on FPGA
title_short Comparison of Vitis-AI and FINN for implementing convolutional neural networks on FPGA
title_full Comparison of Vitis-AI and FINN for implementing convolutional neural networks on FPGA
title_fullStr Comparison of Vitis-AI and FINN for implementing convolutional neural networks on FPGA
title_full_unstemmed Comparison of Vitis-AI and FINN for implementing convolutional neural networks on FPGA
title_sort comparison of vitis-ai and finn for implementing convolutional neural networks on fpga
publisher FIUBA
publishDate 2024
url https://elektron.fi.uba.ar/elektron/article/view/200
https://repositoriouba.sisbi.uba.ar/gsdl/cgi-bin/library.cgi?a=d&c=elektron&d=200_oai
work_keys_str_mv AT urbanopintosnicolas comparisonofvitisaiandfinnforimplementingconvolutionalneuralnetworksonfpga
AT lacomihector comparisonofvitisaiandfinnforimplementingconvolutionalneuralnetworksonfpga
AT lavoratomario comparisonofvitisaiandfinnforimplementingconvolutionalneuralnetworksonfpga
AT urbanopintosnicolas comparaciondevitisaiyfinnparaimplementarredesneuronalesconvolucionalesenfpga
AT lacomihector comparaciondevitisaiyfinnparaimplementarredesneuronalesconvolucionalesenfpga
AT lavoratomario comparaciondevitisaiyfinnparaimplementarredesneuronalesconvolucionalesenfpga
_version_ 1859522237987553280

Ejemplares similares