Small numerical variations between NCS1 and NCS2

[julien_cl]

Hi,

I'm currently testing OpenVino with NCS2

I'm following those instructions to compile squeezenet1.1 and I'm obtaining slightly different results between CPU,  NCS1 and NCS2 devices. 

Using the classification_sample application : 

NCS1 :

classid probability
------- -----------
  817    0.8422852
  511    0.0915527
  479    0.0393982
  751    0.0093536
  436    0.0068550
  656    0.0036659
  586    0.0023270
  717    0.0015497
  864    0.0010500
  581    0.0005631

NCS2 :

classid probability
------- -----------
  817    0.8295898
  511    0.0961304
  479    0.0439453
  751    0.0101318
  436    0.0074234
  656    0.0042267
  586    0.0029869
  717    0.0018148
  864    0.0013924
  581    0.0006595

This is not monstrous but significant.

CPU :

classid probability
------- -----------
  817    0.8363345
  511    0.0946488
  479    0.0419131
  751    0.0091071
  436    0.0068161
  656    0.0037564
  586    0.0025741
  717    0.0016069
  864    0.0012027
  581    0.0005882

I've tested on different networks and globally I have the sensation that NCS1 and CPU are quite similar, but NCS2 is significantly different. 

Is it possible to have access to the error margin we should expect using NCS1 and NCS2 vs the original model before OpenVino conversion ?

Best,

[Shubha_R_Intel]

Dear julien_cl,

I'm sure you recognize that NCS1 and NCS2 are totally different hardware. OpenVino developers do not have this error margin data [for every single network and topology] readily handy to be able to produce for you.  While we still support NCS1, NCS2 is the platform that we are recommending customers use for new applications. I'm afraid that experimentation (as you've done) is the only way to obtain such data.

Thanks for your patience and thank you for using OpenVino !

Shubha