Updating hardware using Allwinner V5?

Question

Have you considered making updated hardware based on the Allwinner V5?

Development board: http://www.lindeni.org/lindenis-v5.html

The V5 SOC is under $10 and it supports dual 4K cameras with hardware binocular depth mapping.  It also has a hardware HAAR feature classifier.   The SDK is on github: https://github.com/lindenis-org  The CPU is same quadcore Cortex-A7 as the A33.

I have one of the dev boards and it has been working without issue. The only thing missing is the tool for using the HAAR classifier with OpenCV. Allwinner has promised this tool and say they will release it once it is working properly. Meanwhile they have provided about a dozen pre-built models.

The SOC also has the hardware to support 8-ch of PDM DMICs to build a microphone array. I have hand-built one and attached it to my dev board. I am able to do DOA calculations using ODAS.  https://github.com/introlab/odas

I also have TVM runing on the board. https://tvm.ai/

Answer 1

Very nice project, thanks for sharing!

That processor is great though I would have preferred more cores and a better GPU.

Lindenis mentions support for tensorflow which is a key element for us, but I could not find any mention that this support is hardware-accelerated using the HAAR accelerator. Can you try to run our tensor flow models (e.g., the quantized mobile nets) on that board and report frame rate? If it is much faster than what we get on our current CPU we will indeed look into that processor in more details.

Comments

The problem is the missing tool for converting models to run on their hardware. Allwinner has promised this tool and they say they are working on it. They are aware it is a critical part of the package. The plan is for you to develop your models using OpenCV HAAR support on a desktop, then when it is ready you run their tool and it converts it to run on the V5.

They supply 8 sample models which do work. Those models are able to process at 30FPS on a scaled down frame. There is a demo that draws boxes around people and faces in real-time. These are OpenCV models that have been converted.

The CPU is identical to the existing Jevois. I can run Imagenet on it in 5 seconds. The smaller quantitized models run in 200-500ms. This is only using NEON, no HAAR acceleration.  (i was able to get Imagenet down to 3.5s using TVM to optimize).

Note that the HAAR Classifers are special purpose hardware, they are not just general FPUs like tensorflow uses to do math.
https://docs.opencv.org/3.3.1/d7/d8b/tutorial_py_face_detection.html

---

Tensorflow is CPU-only. It seems that HAAR is software too. Good for h265, but nothing special for AI...

http://files.lindeni.org/lindenis-v5/datasheets/V5%c2%a0V100%c2%a0Professional%c2%a0HD%c2%a0IP%c2%a0Camera%c2%a0SoC_V1.1.pdf

---

PS And no OpenCL drivers, but for Allwinner it is usual

---

Allwinner has released OpenCV support for the V5 hardware now. It has been checked into the lindenis github. OpenCV uses the HAAR hardware.
https://github.com/lindenis-org

Tensorflow is software only. V5 is same CPU as Jevois so performance is the same.