Tapa Ghosh of Vathys.ai is presenting a new technology that aims at dealing with the one bottelneck few people in the hardware for AI are not focusing on: Data movement. Give this man his funding !
(Unrelated: At LightOn, we solve the data movement in a different fashion)
Petascale Deep Learning on a Single Chip , Tapabrata Ghosh, Vathys.ai
Vathys.ai is a deep learning startup that has been developing a new deep learning processor architecture with the goal of massively improved energy efficiency and performance. The architecture is also designed to be highly scalable, amenable to next generation DL models. Although deep learning processors appear to be the "hot topic" of the day in computer architecture, the majority (we argue all) of such designs incorrectly identify the bottleneck as computation and thus neglect the true culprits in inefficiency; data movement and miscellaneous control flow processor overheads. This talk will cover many of the architectural strategies that the Vathys processor uses to reduce data movement and improve efficiency. The talk will also cover some circuit level innovations and will include a quantitative and qualitative comparison to many DL processor designs, including the Google TPU, demonstrating numerical evidence for massive improvements compared to the TPU and other such processors.
Liked this entry ? subscribe to Nuit Blanche's feed, there's more where that came from. You can also subscribe to Nuit Blanche by Email, explore the Big Picture in Compressive Sensing or the Matrix Factorization Jungle and join the conversations on compressive sensing, advanced matrix factorization and calibration issues on Linkedin.
No comments:
Post a Comment