energyi
Posts: 139
Joined: Tue Mar 24, 2015 9:39 pm

Combine Rpi with other low cost device for massive computing power

Tue May 21, 2019 12:41 am

This article compares different AI scenario performance including using a raspberry pi: https://blog.usejournal.com/google-cora ... 7860b8d87a

Adding the coral usb accelerator to a Rpi seems to be a very low cost, high powered alternative, increasing to 4 TFLOPS. Excellent videos from Google I/O 2019.

I have many Rpi's, just ordered a jetson nano but now wondering if just adding the coral usb accelerator is the right way to go. Playing with the idea of AI for radio telescope analysis for finding pulsars.
energyi

bertlea
Posts: 265
Joined: Wed Dec 07, 2016 6:33 am
Location: Hong Kong

Re: Combine Rpi with other low cost device for massive computing power

Tue May 21, 2019 1:42 am

Current Raspberry Pi models only have USB2 and the article in your link mentioned why Pi didn’t work well with Coral Edge USB as the performance bottleneck is more about data rate. If you want to have many low-cost devices that can effectively use Coral Edge USB, then you should consider SBCs with USB3 such as Rock64.

energyi
Posts: 139
Joined: Tue Mar 24, 2015 9:39 pm

Re: Combine Rpi with other low cost device for massive computing power

Tue May 21, 2019 2:17 am

Hadn't thought of other sbc's. Still with only USB2 Rpi is respectable.
energyi

ejolson
Posts: 3084
Joined: Tue Mar 18, 2014 11:47 am

Re: Combine Rpi with other low cost device for massive computing power

Tue May 21, 2019 8:02 am

energyi wrote:
Tue May 21, 2019 2:17 am
Hadn't thought of other sbc's. Still with only USB2 Rpi is respectable.
From what I can tell, the Coral TPU can only be used to run an inference model, not train one. On the other hand, GPU hardware is much more general and can be used for training with back propagation as well as many other things. Though 3 times more expensive than Raspberry Pi, the NVIDIA Jetson Nano appears well suited for small form-factor half and single-precision GPU computing.

In order to efficiently and easily use a GPU accelerator, a high-speed interconnect to the CPU is important. The CPU and the GPU on the Jetson Nano share memory with a bandwidth of 25 GB/s. For comparison, PCI Express 3.0 x16 runs at 15.75 GB/s while NVLink 2.0 runs at 150 GB/s. Unfortunately, USB2 is 0.06 GB/s and USB3 at 0.6 GB/s, while faster, it is not fast enough to reach the performance levels typical for GPU computing.

energyi
Posts: 139
Joined: Tue Mar 24, 2015 9:39 pm

Re: Combine Rpi with other low cost device for massive computing power

Wed May 22, 2019 12:15 am

Thanks ejolson. Interesting perspective on training a model and running a model. Just trying to figure out this whole ecosystem for AI Machine Learning. I like the idea of using a Jetson Nano to train, then use a Rpi Coral TPU to run the inference model. Headless Zero W with a rpi camera and Coral TPU still ~$100.
energyi

energyi
Posts: 139
Joined: Tue Mar 24, 2015 9:39 pm

Re: Combine Rpi with other low cost device for massive computing power

Sat May 25, 2019 2:18 am

New $25 hat for Rpi from Seeedstudio that relates to this thread:

https://www.seeedstudio.com/Grove-AI-HA ... -4026.html

I can't find and performance information for comparison to what has already been discussed.

Best,

energyi
energyi

Return to “General discussion”