[xs83]

I think the Thinkstation M920 can house a single slot short GPU in it, something like the T400 (https://www.nvidia.com/content/dam/en-zz/Solutions/design-vi...) would give you the capability to run modern GPU workloads.

It wouldnt touch a data centre but the simple ability to run CUDA (Rapids.ai for example) would mean you could prototype things pretty efficiently all on a local setup and not have to pay for GPU costs.

[justin66]

In keeping with the spirit of the Pi that sparked this discussion (I mean, a little bit), an NVIDIA Jetson AGX Orin Developer Kit with 64GB of RAM is $2000, and the Jetson Orin Nano kit is $500.

[xs83]

Yeah thats ridiculously expensive for a home lab IMO, the thinkcenters can be picked up for a couple of hundred bucks used and a T400 runs about the same. So $400 for an AI capable home bench would run $1200 for a 3 node cluster - I can live with that.

My 5 node Rpi5 Cluster ran $1500 with NVMe Hats and a PoE hat and cant do any GPU work :(

[justin66]

The AGX Orin has over five times as many CUDA cores as a t400, and the dev kit is designed to do pretty much exactly what it sounded like you wanted. The other path does offer the opportunity to wire together a pile of new and used equipment and generate heat with it, so if that’s your idea of fun, I guess the choice is clear.