Forums - DSP less accurate than GPU?

2 posts / 0 new
Last post
DSP less accurate than GPU?
raymond.wang
Join Date: 24 Sep 19
Posts: 1
Posted: Wed, 2020-04-01 09:58

Hi,

We are using SNPE SDK in-house to run object detection inference using our coco-based models.

We noticed that while DSP inference time is definitely lower than GPU, the DSP for some reason gives quirky results. For example, there were several instances where we saw with the exact same test image, GPU returned 3 inference results, but DSP returned 1, using the exact same model, exact same scene, zero changes. Just swapping between the runtimes during runtime.

Not sure what the cause of this is, whether this is expected, or are we doing something incorrect?

Your help with this matter would be greatly appreciated.

Thanks,

Ray

  • Up0
  • Down0
jonghoon
Join Date: 28 Apr 20
Posts: 5
Posted: Mon, 2020-05-11 23:54

I guess it is "Yes"

GPU is use float is 4byte normaly(except FLOAT_16  OR HYBRID)

but DSP is quatized data to uint8, so the accurate is less than GPU

maybe if you want 3 object output, you can down score of confidence threshold(i guess, but not sure)

  • Up0
  • Down0
or Register

Opinions expressed in the content posted here are the personal opinions of the original authors, and do not necessarily reflect those of Qualcomm Incorporated or its subsidiaries (“Qualcomm”). The content is provided for informational purposes only and is not meant to be an endorsement or representation by Qualcomm or any other party. This site may also provide links or references to non-Qualcomm sites and resources. Qualcomm makes no representations, warranties, or other commitments whatsoever about any non-Qualcomm sites or third-party resources that may be referenced, accessible from, or linked to this site.