LISSOM gpgpu applied to com neuroscience

I’ve just released as open source a research project I’ve been working on for some months.
It’s a port to GPUs of the LISSOM model; this model is an extension of SOM (Self Organizing Map) with lateral connections (both excitatory and inhibitory).
The software is for research purpose only, and can be unstable.
It’s divided into a few branches:
-liblissom_cuda: the core engine to simulate the model
-libMPICUDALissom: an interface wrapper around the previous library enabling computation on multiGPU systems and/or clusters of GPU enabled computers
-clusterlissom: a software for quick tests, which also allows streaming data processed by many LISSOM maps in a cluster

The website of the model is: http://homepages.inf.ed.ac.uk/jbednar/rflissom_small.html
The website of the GPU port is: http://lissom.googlecode.com

Speedups are some 9x on single GPU, but scale almost perfectly adding GPUs / using a cluster of GPUs.
To this extend, I’d like to point out that I may be able to simulate a large portion (say, 30%) of human V1 after the last part of the software is completed.

Cheers,
Giacomo Spigler

PS: this post is just for scientific interest purpose. Feel free to try out the program, or to browse the wiki @ googlecode (there is a small, yet detailed, pdf presentation of the project), or whatever you want.

Congratulations!!

9x speedup against what ?? – I mean – speed of CPU and few other details (O2 optimized i would presume) would help

Speedup was benchmarked with GTX280 against official LISSOM simulator, run on an Intel Dual Core @ 2.4Ghz [it was also run on more powerful CPUs by the authors, and time taken was about the same].

The benchmark consisted in running 10000 trainining iterations on a 50x50 LISSOM map, recording the time taken.