uh311ae@sunmanager.lrz-muenchen.de (Henrik Klagges) (04/30/91)
Hello ! I heard that neural networks hardware is advancing. Current specifications lie in the range of: - 5 giga connections / s peak - 1 G conn. updates / s backprop learning rate - 1 M connection capacity - integer representation with 1, 8, 16 bits, signed/unsigned - workstation interface with Ethernet , net algorithm library, debugger, visualizer, C interface Do you think this is real ? Do you know about companies actually selling/using such units ? Specifically, is 8bit quantization enough for most algorithms - this would be nice, because then the logistic acfunc could be put into a table (super-hoopey-fast) ? Last not least, what the hell do we need this for ? Thank you for feedback, yours, Henrik Klagges University of Munich, Physics Department
arms@cs.UAlberta.CA (Bill Armstrong) (05/01/91)
uh311ae@sunmanager.lrz-muenchen.de (Henrik Klagges) writes: >I heard that neural networks hardware is advancing. Current specifications >lie in the range of: >- 5 giga connections / s peak >- 1 G conn. updates / s backprop learning rate >- 1 M connection capacity >- integer representation with 1, 8, 16 bits, signed/unsigned Adaptive Solutions Inc. of Beaverton, Oregon, announced a 64 processor chip that was supposed to do 12.8 Giga connections/second with 1 bit weights for feedforward use. Sorry I don't have more details. -- *************************************************** Prof. William W. Armstrong, Computing Science Dept. University of Alberta; Edmonton, Alberta, Canada T6G 2H1 arms@cs.ualberta.ca Tel(403)492 2374 FAX 492 1071