I’d like to prevent building a completely new server just to run a gpu for AI workloads. Currently, everything is running on my laptop, except it’s dated cpu only really works well for smaller models.
Now I have an nvidia m40, could I possibly get it to work using thunderbolt and an enclosure or something? note: it’s on linux
regarding the pricy enclosures, there are vastly cheaper eGPU solutions especially if you’re able to utilise the on-board M.2 or mini-PCI slot. if you don’t move the laptop around, it’s a viable option. this would be an example - not an endorsment. you’d need a $15 PSU to power the graphics and it works well in linux, with the hotpluggability being the primary issue; if you’re willing to shutdown before attaching the eGPU, close to no issues.
you can run it as graphics card (i.e. utilize its display outputs) or just use the laptop’s display with optionally switching between the onboard and discrete graphics.
The M40 doesn’t have outputs :p
But that adapter looks nice! Thank you!
If you’ve got a thunderbolt port on your laptop and a thunderbolt dock on your laptop then there’s no reason why it shouldn’t work.
I’m not familiar with thunderbolt on linux, but on windows you plug it in and it just works™️ and shows up as if it was inside your machine. Your DE on linux might automatically do it, but if you’re command line only you’ll probably have to run a command first.
I did some more searching, and found that nvme to pci-e adapters are affordable. That’s going to look a bit janky, but fortunately I don’t care.
Thank you again for the suggestion!
worth a shot then, worst case I return the dock thing
I did some quick googling. Are those thunderbolt docks really $350 ? That’s like half the price of a cheap computer?!
That can’t be
googling …
holy damn you’re rightMaybe you should do the maths on other options. You could get a refurbished PC for $350. Or buy the dock anyways. Or spend the money on cloud compute if you’re just occasionally using AI. Idk.
I did not say occasionally. We use AI a lot. Currently it’s mostly for image indexing, recognition, object detection, and audio transcription. But we’re planning to expand to more, and we’d like to use models that are more accurate
Fair enough. Yes I figured you probably wouldn’t have a M40 lying around by accident 😅
Oh it kinda just got here somehow. I don’t know why, all I did was buy the thing
By the time you buy the enclosure you can get another old computer. Pickup a old workstation and put the GPU in it. Be mindful of power requirements
There are external GPU cases that might work with your laptop, but at least on older models these were relatively bandwidth limited which doesn’t matter that much for gaming, but I guess it might cause more problems with AI workloads? On the other hand, maybe not if the model fits completely into the vRAM of the m40?
