Faster neural simulations with FPGAs

Posted by Bayle Shanks at 9:11 PM EST

This paper describes the creation of a set of Matlab scripts that allow a researcher to efficiently program an FPGA to simulate a conductance-based neural network model. The researchers describe the use of their system to run a conductance-based model of 40 neurons with all-to-all connectivity at up to around 8x real-time (I wonder how long it takes to run the same model in software on a typical PC?).

Randall K. Weinstein Michael S. Reid Robert H. Lee. Methodology and Design Flow for Assisted Neural-Model Implementations in FPGAs. Neural Systems and Rehabilitation Engineering, IEEE Transactions on [see also IEEE Trans. on Rehabilitation Engineering]. Volume 15, Issue 1. pp. 83-93. March 2007

FPGAs (Field Programmable Gate Arrays) are hardware with a bunch of logic gates, flip flops, and such, which can be dynamically re-connected. So a neural network model running on an FPGA can be expected to be much faster than one running on a general purpose computer, because the model is effectively being run on specialized hardware designed for that task.

The FPGA hardware they were using is a product similar to this, which is about $2,500! So it’s not very cheap, yet. But we can expect this sort of thing to get cheaper at a geometric rate a la Moore’s Law.

Note that the 8x figure is when the network is running as fast as it can, so fast that the computer can’t reliably capture intermediate results from the simulation. It runs slower if the computer is recording everything; they don’t say how much.

I was hoping to find that the Matlab scripts that they described building would be open-source, however I didn’t find a URL in the website and couldn’t find a lab website online. Also, it seems like they are planning to commercialize this, so they may not release it at all. :(

After this paper, the first author, Randall K. Weinstein, became the co-founder and Chief Systems Architect of http://www.simatratechnologies.com/. A search over NIH awarded grants reveals grant 1R43NS057859-01 for “A low-cost, high-speed platform for neural modeling”, which this page suggests may be worth on the order of $183,163 for 2007 (or maybe that’s the amount that has already been dispensed? or the amount left?). Description of the grant:

Neural models, mathematical descriptions of neural behavior, are an invaluable tool for developing new medical treatments and understanding how the nervous system works. But as researchers discover more information about the nervous system, these models become more complicated. As a result, many modern neural models require powerful computer hardware, such as supercomputers, in order to be simulated and studied. Unfortunately, these systems are expensive and difficult to use. This project will create a low-cost, user-friendly, and computationally powerful system for neural modeling based on a technology called field-programmable gate arrays (FPGAs). This project will develop the user-friendly tools for creating neuron models on FPGAs, and the high-speed interface that will maximize the computational power of FPGAs.

Incidentally, what with the Georgia Tech Laboratory for Neuroengineering and companies like this springing up in Atlanta, seems like Atlanta (and more broadly, the U.S. southeast; the DeMarse lab is in the University of Florida) may be becoming a hotbed of neuroengineering.

One Response to “Faster neural simulations with FPGAs”

  1. Temple University Neural Instrumentation Lab » Large-Scale Neural Simulation Says:

    [...] [...]

Leave a Reply