Finally, "real" GPUs on the Raspberry Pi 5

you can choose super performant GPU and use it with super slow disk storage (SD card) or you can choose super performant disk storage (NVME drive) and use it with slow integrated GPU. Hard decision...

Or you can spend a lot of money on a PCIe switch to share the limited PCIe bus bandwidth between them.

On a more serious note ... sometimes YouTubers are so preoccupied with being first to try something that they forget whether they should do that "something" in the first place.

An expensive GPU deserves better :).

- Gough

misaz said:

you can choose super performant GPU and use it with super slow disk storage (SD card) or you can choose super performant disk storage (NVME drive) and use it with slow integrated GPU

Though you missed a couple of options, no?

1. Network storage.

2. USB3 storage.

Both would be faster than SDCard. In fact, perhaps you could cluster a number of them together to take advantage of the compute, arguably there may be cheaper ways to do that.

And here I was thinking it was about providing CUDA cores to be used by userspace programs for inference tasks...

Eh, rather a Jetson Nano or relative - a lot neater and still quite compact. But perhaps not as much firepower ... or as much plug-melty action.

- Gough

Network storage has horrible latencies and is almost unusable for serious gaming. USB 3.0 has similar issue in theory, but not so significant. With USB 3.0 SSD it may work, I think. But still can't compete with x86 solutions.

Gough Lui said:

Eh, rather a Jetson Nano or relative

Are many people using a Jetson Nano? I don't see many people posting about them on the Community.

More than you'd imagine, I'd say.

We have a good slice of our PhD candidates using them as targets for AI in embedded contexts for computer vision especially. They really fly compared to RPis for that particular use case. Hook them up to some RPi cameras (compatible CSI interface) for some fun :).

Some have even invested into the Jetson AGX or clusters of them. There are some autonomous vehicle researchers using them too - likely as things would be easier to port to Nvidia's DRIVE platform.

EDIT: Perhaps why we don't hear much here is that most of the people I've seen working with them are mostly focusing on the software, rather than the electronics side. Everything to them is code ... and this isn't the "forum" they would naturally gravitate towards. That's not to say the board isn't great for plenty of electronically-related shenanigans too.

- Gough

Right now there is a lot of movement around the  the NVIDIA Jetson on hackster.io, 

AI Innovation Challenge - Hackster.io

Several regular members of element14 participating in it.

I'm using it . Still learning how to use that (with the NVIDIA specific functions - the Jetson is basically a computer that runs Linux) stuff, but it's better and faster than a Raspberry PI.

I have the Jetson Nano developer KIT and it's not the most up to date stuff. Python is 3.6 and because of that I'm limited on the TensorFlow, Pytorch versions that I can install.

Because I still have to receive my PI5 (4GB) can't compare them yet, but as soon as I get my RPI5, I will definitely compare them.

Regarding the original post, my question goes - Why ? Why run an external GPU ? If I have the need to run something that requires a GPU, I will not use a Raspberry PI.  A Jetson Orin or something else that is specific for that task.