I want to connect a GPU GT640 to a Raspberry pi model B. I know there is not a direct connection and the tranfer data are very differents but. Is there any chance to connect this devices in order to take advantage of GPU processing power? is not necessary to have a real time communication, but is this possible in theory?
It would be possible, but please be aware that you will need to write your own drivers and sort out the power issues, because I doubt the RasPi will be able to power a full PCI/PCI-E GPU on it's own. Please also bare in mind that the CPU will also affect the GPU's power, so it may be an idea to research before you try. <--- Adding to that, you can use a 1x to 16x powered riser cable commonly used in gpu mining rigs. There are other powered riser cables that utilize adapter cards for USB as well.
You could use a startKIT which has a PCI Express x1 card slot which you could use to interface to a suitable GPU (you would likely have to write your own drivers), and this could then interface with the Pi though the GPIO header.
So in short: yes you can, all the hardware to do so exists and is readily available, all you have to do is write glue logic, but don't expect it to be easy.
So, to wake the dead thread, I would not do it on a Raspberry Pi personally. I would recommend at least the Pi 3 with a mPCIe to PCIe adapter designed for desktop GPU and powered by a dedicated power supply. Also you need to make sure that the version of Linux has driver support for it; if there's no drivers someone has to write them, and that can take months with the right assets and source code. It is 100% possible to do this way, but you will likely have to be talented enough to write it for the community if it has not been done already.
Just bumping to an old thread. It's nearly impossible, but the main reason is capitalism and bureaucracy, not writing drivers.
Sure, writing drivers is very difficult, but work with an epic team and it would be done before sunrise.
Even harder would be to get access to existing source codes for the whole video card and maybe even source codes for either the BCM2835, BCM2836, or BCM2837 (Pi 1, 2, and 3 respectively). You need those to learn how the cards work on the inside and find a way to interface with it with each other.
You would have to be an OEM to get just even a chance to get the source codes. Even the Raspberry Pi foundation itself
has had has difficulty with those and has to settle on a pre-compiled blob to interface with the processor and GPU.
Power requirements wouldn't be an issue at all since, by the time you're begging [company name], you would have already figured it out. Probably a PC PSU connected to the video card and also powering the Pi via its 5V rail.
Your only problem will be drivers. It would be so cool to get a Nvidia GPU to work on a Pi, but this is almost as hard as building your own GPU from scratch.
People are right, the hardware isn't impossible, but it WILL be sketchy. You absolutely will need drivers, however. Even if you could write your own, they would not be optimized to utilize all to the speed and features on the GPU.
I wanted to try this with an old Nvidia GForce, but I am a lousy programmer with no hardware hacking experience.
I figured it was a long shot, but I emailed NVidia just after the RTX was released, and asked if they would either compile the drivers for a few Pre-2010 GPUs for the Pi and make them available to download and tinker with, or release the source code for people to compile themselves. Obviously they said "no."
So here is the thing, we have external GPUs for laptops (costs disregarded) and the Wi-Fi in the new Pi 3B+ is supposed to be fast, why not see if that could work?