Trying to Get an Nvidia GPU to Work with Raspberry Pi CM4 | Hacker News Reader

Trying to Get an Nvidia GPU to Work with Raspberry Pi CM4

gist.github.com

22 pointsgeerlingguy6 comments

I'm posting here for two reasons:

First, to show that you can plug in a GPU (in this case an older one, but still passable) and have it recognized by the Pi Compute Module 4.

Second, to mention that it seems the BAR (Base Address Register) space allocated to PCIe on the Pi 4 is not adequate for even an older GPU like this one.

Some people have suggested some GPUs may work but with much reduced bandwidth due to the x1 lane on the PCIe 2.0 bus on the Pi 4, but so far from all the fruitless searches I've done, anyone using any of the Cortex A series chips seems to be bumping into the same problem (BAR memory allocation problems due to lack of space) and nobody else seems to have found a solution either :(

Anyways, it would be interesting to know if this is a hard limit of the Cortex A72, or something else, and if something else, whether that limit could be lifted (e.g. firmware update or OS changes).

Why would it be related to the Cortex A72? It's just going to have a bus coming off it, right? It's got no idea what PCIe even is. The problematic bit would seem to be the PCIe controller/root complex on the SoC.

I doubt it's a limit of the CPU. It might be a limit of the method used to connect the PCIe component to the CPU (the SCB bus on the PI IIRC).

The device tree file does contain the range of memory to use for the PCI BAR, could it be as simple as increasing that number (the last value of the 'ranges' item I think) and rebuilding it? Seems unlikely, but might be worth a go. I'm uncertain how much the device tree file is describing the actual setup of hardware, and how much is used to actually configure the hardware.

The larger issue might be the complete lack of I/O space. Again this could be added in the device tree file, but who knows if the underlying device even supports it.

Well that sucks. I suppose it makes sense that you might run into an issue like that on an embedded board, but I had hoped it would be simpler.

This means that you may have issues with some of the higher end HBA's, etc. as well. I'd have to do some research to see what they require though.

Since the AMD Navi Driver are in Linux distro wouldn't it be better try those cards first?

I don't have an AMD GPU to test with, any cheap recommendations? Especially if they are slightly older and maybe have only 1-2 GB of RAM, since it seems larger/newer cards won't be able to work with the Pi's PCIe bus.

Reply on news.ycombinator.com