Optically connected and reconfigurable GPU architecture for optimized peer-to-peer access

Erik Anderson, Jorge González, Alexander Gazman, Rodolfo Azevedo, Keren Bergman

Producción científica: Capítulo del libro/informe/acta de congresoContribución a la conferenciarevisión exhaustiva

Resumen

Increasing industry interest in the optimization of inter-GPU communication has motivated this work to explore new ways to enable peer-to-peer access. Specifically, this paper investigates how reconfigurable optical links between GPUs in multi-GPU servers can allow for minimized memory transfer latencies for given machine learning applications. Silicon photonics (SiP) is proposed as the enabling technology for such a reconfigurable architecture due to the potential for scalable and cost-efficient production. We evaluated our architecture using traffic obtained from an NVLink-connected 8 GPU server executing a set of machine learning models including AlexNet, DenseNet, NASNet, ResNet, MobileNet, and VGG16. Our results show up to 24.91% reduction of the total relative transmission latency (RTL) between peers.

Idioma originalInglés
Título de la publicación alojadaMEMSYS 2018 - Proceedings of the International Symposium on Memory Systems
EditorialAssociation for Computing Machinery
ISBN (versión digital)9781450364751
DOI
EstadoPublicada - 1 oct. 2018
Publicado de forma externa
Evento2018 International Symposium on Memory Systems, MEMSYS 2018 - Alexandria, Estados Unidos
Duración: 1 oct. 20184 oct. 2018

Serie de la publicación

NombreACM International Conference Proceeding Series

Conferencia

Conferencia2018 International Symposium on Memory Systems, MEMSYS 2018
País/TerritorioEstados Unidos
CiudadAlexandria
Período1/10/184/10/18

Huella

Profundice en los temas de investigación de 'Optically connected and reconfigurable GPU architecture for optimized peer-to-peer access'. En conjunto forman una huella única.

Citar esto