The Raspberry Pi has a videocore 4 GPU on it. That is what I was trying to use. Its not an open hardware piece so docs are sparse. I found an article on how to use it for FFT's in the shaders which is what I want to do but trying to implement it proved to be problematic and its nowhere near as fast as the K1. Has 4 cores.
So I'm going to use the PI's for other things. One possibility is a Hadoop cluster which is used for big data sets. Like a bunch of images. One of the boards I took out of the original stack is now running my backups. Its running freeNAS and is connected to a small raid array. The only truly open gpu is on the beagleboard black. But again no where near the power of the nvidia chips. Nvidia has the best software support and tools as well. All kind of libraries for the Tegra. OpenCV which is the standard for computer vision applications. Cufft lib for doing fft's. Several other math orientated libraries. OpenMPI for clustering with cuda support.CuDNN deep neural network libs for maching learning. Used in self driving bots. The little beast is doing some pretty awesome things. The automotive version has two tegras on the same board. Check out some of the stuff they are doing with it here.
http://www.nvidia.com/object/drive-px.html The computer used to do this is basically two TK1's on a single card. Their site is full of amazing things people are doing with GPU hardware. Low power 12v operation compactness. Its going in a box sitting at the telescope. I'll communicate with it via X-windows over wifi. Its going to house all the programs for moving the scope shooting images etc too. And I'm going to use it on another robot project I have been thinking about doing for some time now. But to answer your question the older hardware is fine I have an older board I sometimes put in with the 660TI for more cores. There are some limitations as to how far you can go back.
On the PI side each one will have a specific task and the rest of the cores will be in the pool to be used as needed. I have the wireless touchscreen handpad working now good to 100m and the dew heaters and sensors for that. Just the focus left to do and I want to use a variation of Speckle interferometry to do that. Also you can use FFT's just like you do FWHM for determining focus. The two outputs will come closer as you get closer to focus and widen up as you go away.
I wouldn't say its for everyone the admin tasks alone are enough to scare most people off but if someone had a distribution they could download to one it would be much easier. Just getting all the programs downloaded and compiled has taken a couple of months. I can't put in a lot of hours a day on it and some days I just don't have the energy. So someone a little younger with more energy could probably whip it together in a week.
I posted up some pics of it on another forum and ran into the guys doing the Speckle. I didn't know about it just through getting on the Mathematica site and looking for ways to do FWHM I ran across using FFT's to determine sharpness.
I went with the PI's at first because they come with Mathematica on them. It has a ton of functions that directly relate to astronomy. You can move your scope and image with it if you want. I found a tutorial that has some killer functions to deal with images. One will split out every star in an image into a separate image for instance. Once they are in Mathematica they are already in an array you can operate on. Not being a math whiz I was leveraging my meager skills with the software. But to do a burst of images in a short enough time not to get in the way just required a lot more computing power than the PI's had to offer.
I bought a notebook less than a year ago and the dew has already messed up the keyboard. So I want something that I can put in a weatherproof box and not worry about it. I live right across the street from a lake its always dewy here.
So that is basically the story of how it all came about. Now I have some people to lean on for the math and I'm benchmarking the Jetson so they can decide if its right for their project in return. Just a couple more things to set-up and they will be able to log in and work on their stuff as well. They are 1024x1024 2D kernels so anything over 1024 cores will be optimum so you can run one element of the calculation on each core. Say it takes 10ns for a calculation and you are doing 1024 of them that is 10240ns for the total. With the GPU and enough cores it will take exactly 10ns for the whole thing. The test FFT I've been running is a 2048x2048 2D using 32 threads. Just using the sample programs I haven't even started to read up on the code. Fortunatly nvidia has a 6 week course on their site free that will teach you parallel programming with some interesting projects along the way in image processing. So I've been going through those to pick up the details. So far its a very nice course. I've only gotten through 2 lessons so far. The TK's came in Sat. Took all weekend to get the linux set up on my pc and flash both TK's and get them set-up. Linux is new to me I'm a BSD kinda guy. Linux was just getting popular when I quit working. But its starting to make sense now. Lots of new things new ways to configure networking displays etc.
The pics are the three screens on the handpad. It has a config screen as well. The sensors and handpad and the Mathematica program figuring out the centrioids in a star field. Its worth the cost of a PI just to get Mathematica in my opinion. Just make sure you get the PI-2 with the 4 core cpu. However there are free programs like scilab that are almost as complete you can run on any box.
Griz
Found the Pi's for 22bucks apiece. 12 for those Drok power supplies. It also functions as a power amp and volt meter.
