topic Questions about parallelism and data type in NCS in Intel® Distribution of OpenVINO™ Toolkit

Questions about parallelism and data type in NCS

idata — Thu, 21 Dec 2017 16:43:33 GMT

Hi all,

I've tired to use the stick these days and encountered 2 questions.

As far as I know, NCS receive high computation speed when the input images and models are large, but when I input a small image(like 24*24) to a light model, the computation capacity of the stick are not fully utilized. So my question is can we input 2 or more images in parallel and assign different SHAVEs to process them? Or is there any better way to accelerate light model?

I would like to know the data type used in the stick, is it 16bit fixed-point or not?

Thanks in advance.

Andy

Re: Questions about parallelism and data type in NCS

idata — Fri, 22 Dec 2017 00:13:14 GMT

@xuzijian418 Thanks for your interest. At the moment, there isn't a way to process images in parallel by utilizing multiple SHAVE processors. The NCS was designed with a balance of power efficiency, accuracy and inference speed in mind and as such the data type used in the NCS is float16.

Re: Questions about parallelism and data type in NCS

idata — Fri, 22 Dec 2017 09:43:42 GMT

@Tome_at_Intel Thank you Tome for your nice answer! i'm looking forward to seeing your new powerful product :)