neural netwrkng: optimze training bfore fltering complxity

Moderators: mdrjr, odroid

neural netwrkng: optimze training bfore fltering complxity

Unread postby back2future » Sat May 05, 2018 8:53 pm

seems to be a growing market by 1:4 for next 3 years

machine learning for consumer markets ... neural networking ip's need trained dev frameworks and libs before being useful for performance and consumption improvements
https://en.wikipedia.org/wiki/Convoluti ... al_network

[ xu4, tensorflow, near realtime object_identification 2017: viewtopic.php?f=95&t=28177 ]
naturally beYOnd spectrum
back2future
 
Posts: 75
Joined: Sun Jul 23, 2017 3:19 pm
languages_spoken: english

Re: neural netwrkng: optimze training bfore fltering complxi

Unread postby ASword » Wed May 30, 2018 1:04 am

The upcoming generation of SoCs appear to have quite a few with dedicated "neural processing" engines. The RK3399-AI has an "NPU" (not to be confused with _network_ processing unit) rated at 2.4 TOPS and very low power draw -- quite amazing. Only suited to inference, however, so training will need to continue to be done elsewhere (i.e. a more powerful machine, or else on the GPU/CPU of the SoC) and in the background while inferencing runs in realtime.
ASword
 
Posts: 76
Joined: Fri Aug 04, 2017 12:48 pm
languages_spoken: english
ODROIDs: XU4, HC1


Return to General Chat

Who is online

Users browsing this forum: No registered users and 1 guest