Has anyone experience of running distributed AI on the Cluster Hat using Pi Zeroes? Specifically, can I run TensorFlow on a Cluster HAT ideally using the GPUs on the PiZeros as well as on the controller?
Well, you reminded me about one year ago I was very excited to hear that Rpi Zero can do Google TensorFlow. So I bought 3, one Zero and two Zero Wifi to try out.
But then I heard one very bad news - the Rpi GPU cannot do TensorFlow, because of license problem or something. So I switched to Nivida GeForce GPU.
I later heard one more bad news - the RpiZ communication channel is very slow, much time is wasted in data traffic, rather than computation. So RpiZero TF was a bad dream to me.
Besides, when I first tested out RpiZ and RpiZW, I was very disappointed, because everything goes so slowly, compared to Rpi3B. So I just gave away the 3 RpiZs to my 12 year old nephew.
I also read that Google was going to release an Rpi compatible Edge TPU (Tensor Processing Unit) Accelerator USB stick, which they claim runs faster than the Movidius Neural Compute Stick
One good news for Rpi fans is that Google has recently switched from Ubantu to Debian, but one bad news is that there is still the ridiculously slow Rpi USB2.0 bottleneck.
You also reminded me that I actually signed up for the stick which they said would came out 2108 Fall. I have not heard any good news from them. I wonder if I would wait past 2019 Fall.
PS - My bad experience is with Rpi Zero W V1.1 2017, and Rpi Zero V1.3 2015. I am not sure if newer versions have improved. I attach my test report below.
My disappointing Rpi Zero Experience - tlfong01 2018aug01
Rpi Zero V1.3 2015 is very disappointing, will switch to Rpi 3B+
I have setup the Rpi Zero V1.3 2015. The results is very disappointing.
Rpi cannot recognize my USB3, switchable USB hub. I need to use a cheap, not so reliable hub instead.
The screen resolution is very low. I guess it is TV mode or similar 640 x 480 resolution (Rpi Zero W did not have this problem, and automatically recognizes my high resolution 1280 x 1024 monitor). I googled and found people talking about how to set the resolution. I have the feeling that it is a complicated and time consuming thing to do.
I already found Rpi Zero W V1.1 2017 a bit slow (disappointingly slow when using Geany's GUI commands). This Zero V1.3 2015 seems slower, and more stupid. I think it is not time efficient to use it for hardware/software development. So I decided to give up, and switch back to Rpi3B+, ...
The following article is a very good update. So now I know which old AI toys disappeared, and the new thing - Google Coral comes out.