Ticket Name: Deep Learning and Inference accelerated on DSPs

Query Text:
Other Parts Discussed in Thread: TDA2 Hi, With the multitude of DSPs built into the processors from J6 -> Sitara -> Multicore ARM+DSP, is TI working on or already supporting Deep Learning inference acceleration support? Specially as it applies to applications very similar to the ADAS and autonomous navigation applications targeted by the Jacinto series?

Responses:
Hi Anup, I will ask our experts to comment on the keystone device support on "Deep Learning inference acceleration" and road map if any. For Jacinto, I knew it is for automotive devices. Can you please post it on Automotive forum for appropriate and faster response. Thank you.

Hi Anoop, We have the vision library - VLIB 3.3 ( software-dl.ti.com/.../index_FDS.html) supporting some of the important kernels of deep learning (convolution, maxpooling, Fully connected) - please refer the library and documentation. This library is for C6x DSP and can be used on any TI device with C6x DSP. We have on our roadmap to add more functions for deep learning and appropriate framework support. On Jacinto family (specifically ADAS parts like TDA2/3x) we offer this as part of vision SDK. Please get in touch with TI field representative to get access to vision SDK. Thanks, With Regards, Pramod

Hi Pramod, Thanks a lot for the pointers. We will check out the VLIB, and get in touch with the local FAE at an appropriate juncture. Wish to see a lot more proofs of concepts to add confidence on TI DSPs though! :)

I wouldn't hold my breath ... the C66x architecture is quite old by now. It does not have enough processing power for most deep learning algorithms. -Robby

HI Robby, TI SOCs have hetrogenous processors (EVE and DSP) and multiple instances of those which gives a good horse power for many applications using deep learning technologies. Stay tuned to see some cool demos on this technology at CES - Jan 2017 @ TI booth Thanks, With Regards, Pramod

That's good to know. I'd be interested to seeing those demos. -Robby

Dear Champs, TIDL (TI deep learining library) was introduced in CES2017 (Youtube link below). Could you share more info about TIDL? Is there a landing web page for TIDL?

Hi Luke, We are planning to release it and you should see more documentation and information with that. The releases will be during 1Q and 2Q this year. Thanks, with Regards, Pramod

Hello Pramod, Is the TI Deep Learning library available for usage? I am planning to use this on J6 (DRA75x) EVM running Android. Can you please let me know if this is possible? From the video of TI demo at CES 2017, it seems that the demo was on TDA2 and not on J6 but from the spec of TDA2 it seems similar to J6. Regards, Pavan D

HI Pavan, On J6 we have 2xEVE and this demo was done using 4xEVE on TDA2x. You can run the demo on J6 with half the speed. Thanks, with Regards, Pramod