11-07-2012 05:23 PM
I have and HP DL585 G2 machines ... she has 8 cores and 65GB of RAM and 6 free ecpi slots.
What I would like to accomplish is build a low-end (primarily for learning purpose) a Cray XD1 like machine.
From my limited understanding a Cray XD1 was an x86 based platform that had a couple of FPGA boards that could be programmed to do custom computation.
My design parameters are:
1. Low cost -- buy FGPA boards off of ebay or the like
2. use LINUX as my general purpose computing platform
3. preferrably use LINUX and open source based tools to design the circuits that I would download onto the boards.
4. Again this main purpose is teaching/learning.
5. My primary areas of applicaiton are scientific computing: matrix arithmetic, vector computing.
1. Any serious problems you see with what I am trying to do?
2. What board should I be looking at or other hardware?
3. What software should I be looking to work with that board?
4. Any books that your recommend for doing what I'm trying to do
Thanks in advance.
11-08-2012 12:24 AM - edited 11-08-2012 10:57 PM
seems like you are planning to do sth. like this:
You should find out a little more details about the Cray XD1.
- How were the FPGAs connected to the CPUs to ensure performant data transfer.
- Did the use the PowerPC cores in the V2pro devices?
- What kind of operations were actually supported by the FPGAs? (Reconfigurable?)
There are many actual boards that can easily be connected via some PCIe slot, so connecting them to your host is no problem.
But there are so many ways that you can use the FPGA then...
You need to specify the operations and how the FPGA and CPU should work together.
This can become very complex.
Have a nice synthesis
11-08-2012 08:30 AM
Hi Many thanks for the response.
I'll look at the link you provided.
So just as a clarification I was not trying to replicate the Cray machine 1 for 1.
But rather, the idea I had was say that I want to do some large matrix multiplication with say big number precision I could reconfigure my FPGA boards to do that and then pipe my data from my LINUX machine to the FPGA boards and then out again.
So an anlagous system that one could compare this to (from the point of a noob like me is) when I add say 6 GPUs to my LINUX machine and use CUDA to speed up special tasks.
11-08-2012 09:30 PM
Depending on the kinds of computations you do, you might be interested in Xilinx System Generator, which allows you to generate FPGA designs using Simulink. If you're already experienced with Matlab/Simulink, this is supposed to provide an easy way to provide hardware acceleration.
11-09-2012 11:06 PM
Many thanks for the notes -- really very useful.
I guess my next question is that type of boards that I can pick up from ebay for a very reasonable cost should I look @ that support a ePCI interface and onces that I can program with open source tools as I figure this out more.