10 maximum executors on Intel Neural Compute Stick 2?

[AGaru]

NCSAPI states that my NCS2 has 10 maximum executors (which is NC_RO_DEVICE_MAX_GRAPH_NUM, but OpenVINO treats it as maximum number of executors).

- Where can I read about that limitation?

- Why is it exactly 10?

- Can I increase that number for one stick?

Thanks!

Env: Windows 10, OpenVINO 2019 R1.1, NCS2, firmware MvNCAPI-ma2480.mvcmd.

[Shubha_R_Intel]

Dear Garustovich, Alexander,

Where are you reading that NC_RO_DEVICE_MAX_GRAPH_NUM  or _executors  or _maxExecutors is capped at 10 ? I am not seeing 10 in any of the code below (when I do a grep search):

https://github.com/opencv/dldt/tree/2019/inference-engine/src/vpu/myriad_plugin

Also note that we just released OpenVino 2019R2. It has many bug fixes and improvements. In fact one of the new features is Custom Layers for NCS2 (preview mode).

Please download R2 and give it a try. Also, kindly report your findings on this forum.

Sincerely,

Shubha

[AGaru]

Hi, thank you for your effort!

The cap comes from the device to the field max_graphs of type deviceCapabilities_t. You can find it in ncCommPrivate.h.

This field among with others is filled in function getDevAttributes (in file mvnc_api.c). Then later OpenVINO takes that value and store in DeviceDesc::_maxExecutors (file myriad_executor.h).

[Shubha_R_Intel]

Dear Garustovich, Alexander,

OK you mean https://github.com/opencv/dldt/blob/2019/inference-engine/thirdparty/movidius/mvnc/include/ncCommPrivate.h

I have inquired about this question on your behalf.  We should get an answer soon. Thanks for your patience ! Also have you tried OpenVino 2019R2 yet ? Please do ! It was released yesterday. Custom Layers are now available for NCS2 in Preview Model. 

Thanks,

Shubha

[Shubha_R_Intel]

Dear Garustovich, Alexander,

The 10 graphs (or 10 models) limit is there because the firmware supports max of 10. The firmware supports max of 10 because it is a good number carefully chosen two or so years ago. We have ideas about removing this limitation but  right now it is a low priority feature.

Hope it helps !

Thanks,

Shubha

[AGaru]

Thanks!