FriendlyELEC-Forum

Hi,

We are running Face identification on Neno PC T4. We are not getting performance as expected. Below are the result.
Neno PC-T4:
Face Identification on Cortex A72 core: It takes approx 940 msec.
Face Identification on Cortex A53 core: It takes approx 2500msec.

Neno PC-T3:
Face Identification : It takes approx 400 mSec.

We have kept all the parameter same while comparing performance. We have used "rk3399-sd-lubuntu-desktop-xenial-4.4-armhf-20180921" os.

Why we are getting less performance then Neno PC-T3?

I have searched and found that in RK3399 with Linux kernel 4.4 there is DDR performance issue. Is it correct? Is there patch available?

The NanoPC T3 is faster then the T4.
The T3 had got 8 cores at 1.4Ghz (11.2Ghz total)
The T4 has 2x1.8Ghz + 4x1.4Ghz(9.2Ghz total)

So multicore performance is lower on the T4/M4/Neo4 then on T3+.
The advantages of the T4/M4 are the USB3, high single core performance,
Last week I made a video about benchmarking SBC's. There I show the T3+ is the fastest of all my sbc's in multi-core tasks.

https://www.youtube.com/watch?v=EZMHo3bVnOo&t=187s

yashavadan wrote:
Face Identification on Cortex A72 core: It takes approx 940 msec.
Face Identification on Cortex A53 core: It takes approx 2500msec.

difference between 2xA72 and 4xA53 looks a bit too high for me. as well as between 4xA53 and 8xA53 in T3.

did you check any other tests? (from OpenCV, for example, which include faceDetect algorithms with both Image processing and NN )?

yashavadan wrote:
The T4 has 2x1.8Ghz + 4x1.4Ghz(9.2Ghz total)

why you thinsk, that core's in T4 have same A53 architecture ?

I agree with ayaromenok. Did you check to see you were linked to the correct libraries ? Maybe not taking full advantage of the
VPU and GPU ? The very first test I did with my board is install the Friendly Desktop and compile my object tracker code for a robot I built with a TinkerBoard and Core2-ROS board. It was amazed how fast it was. I can't wait to install ROS and all my apps over to this board. Using an nvme for ROS makes a HUGE difference!

Double check to see what your code is linked to. Here's the debugging tool I use.. It makes debugging Mali GPU performance much easier , and it's FREE!

https://developer.arm.com/products/soft ... s-debugger

ayaromenok wrote:

yashavadan wrote:
Face Identification on Cortex A72 core: It takes approx 940 msec.
Face Identification on Cortex A53 core: It takes approx 2500msec.

difference between 2xA72 and 4xA53 looks a bit too high for me. as well as between 4xA53 and 8xA53 in T3.

did you check any other tests? (from OpenCV, for example, which include faceDetect algorithms with both Image processing and NN )? Benefits of Male Extra for Men by Emilydavinci website

yashavadan wrote:
The T4 has 2x1.8Ghz + 4x1.4Ghz(9.2Ghz total)

why you thinsk, that core's in T4 have same A53 architecture ?

It can turn the effects light because of those amateur nature are hitting the cortex A72 Soon.. a real soon