search

Guzunty / Pi

This repository contains resources to support the Guzunty Pi IO expansion board
110 stars 32 forks source link

Guzunty Core For Neural Network Programming #18

Closed dthomsonit closed 10 years ago

dthomsonit commented 10 years ago

Is it possible to use the Guzunty For Neural Network Programming.Does anyone have any experience in this area.Is there any cores available or guidance in developing a neural network core for the Guzunty.Kind Regards David dthomsonit@googlemail.com

campbellsan commented 10 years ago

I have been interested in neural nets for 40 years!

Guzunty is not especially suited to implementing neural nets because nets require to store a lot of state. State == memory, and that is the one thing that a CPLD is not great at providing. Each bit of memory requires one flip-flop and there is one flip-flop per macrocell in the CPLD. That means that the absolute most we can store on board the chip is 9 bytes!

Paired up with a suitable RAM device, it's possible. The advantage of using a CPLD is that the neural net response time would be quite fast compared to a Von Neuman based implementation.

Can you say a bit more about your application? Or was your question a more general one?

dthomsonit commented 10 years ago

Hi campbellsan, Thanks for getting back to me with your experience and assistance on Guzunty.My intentions for Guzunty was for a possible cheaper alternative to the following Cognimem CM1K Neural Network Chip http://www.cognimem.com/ but as you indicated it is not suitable for this type of work.Maybe with some research I could interface the chip that sells for $149.00 http://www.cognimem.com/products/design-resources/CM1K-PGA69/index.html with Guzunty.I will however have to look into the additional electronic components required to interface to the chip then Guzunty. The SDK Development Board is very expensive at $1500.My main aim is to develop simple Neural Network Applications in the field of Computational Neuroscience (Applications: Neurological Disease) using C++ that would run on the CM1K via programs sent from the Raspberry Pi, Kind Regards David

campbellsan commented 10 years ago

Most interesting,

It doesn't look like the module would require much in the way of support components, just a prototyping board and there are plenty of those for the Pi. The schematic in the module hardware manual is not very clear, but it looks like it has been designed to make interfacing with a 3.3v based host very easy. From a quick look at the specs, you can use the I2C interface to communicate with the device. A Guzunty might not be needed, although there are quite a few other signals required, so I'd need to look more closely to be sure there is enough IO.

As you observed, the marketing strategy on the SDK is incomprehensible, and rather Windows centric in some cases it appears. I think getting interface libraries working on the Pi is going to be a bigger challenge than connecting the hardware.

It would be interesting to see what could be done with this device in combination with the Pi camera. You are aware of the OpenCV work being done on the Pi, right?

campbellsan commented 10 years ago

Closing this issue as part of a site cleanup. As interesting as it is, there has been no activity on it for more than 6 months.