Der IBM 'Resistive'-Prozessor beschleunigt den Prozess des Trainings neuronaler Netze um das 30.000-fache

iCover! , , Google AlphaGo , . IBM , .

Bild

IBM (Tayfun Gokmen) Resistive Processing Unit (RPU), (Deep Neural Networks, GNN) .

“, RPU , . , , , IoT”, — arxiv.org.

. , FPGA (ASIC). , , . IBM , — (PCM) (RRAM).



Schematics of original weight update rule of Eq.1 performed at each cross-point. (B) Schematics of stochastic update rule of Eq.2 that uses simple AND operation at each cross-point. Pulsing scheme that enables the implementation of stochastic updates rule by RPU devices for © up and (D) down conductance changes.

DNN ( ) 27–2140 , , , , . , , , , , 30 . .

, -. , , , , , «» . IBM , , .

, IBM, T.J. Watson Research Center, , Google IT-, . , , .


(pdf)


, . , , , . , , .



Source: https://habr.com/ru/post/de392285/


All Articles