[bm1362]

I just went this route to build a ML workstation. I was able to get a prebuilt with a 3090 for ~30% of a comparable build on https://lambdalabs.com/.

Now I just need to wait until next year to get the second 3090. I considered waiting for the new A4000s to launch to see if they'd have better inventory but figured 3090s have additional value as a gaming gpu.

I'd be curious how professional ML folks are fairing and what they're using -- it's not too late to cancel my build.

[thebruce87m]

Just got this delivered yesterday:

https://www.novatech.co.uk/savedbuild/ccf0fcec-685d-4f4b-8be...

New to ML but looks similar enough to some of the builds on your site that I’m slightly less scared that I made the wrong choice!

[bm1362]

Dope, 7 days is a pretty nice turnaround. I went with a 'custom' prebuilt so I could plan for adding additional cards once they're well supplied. You might run into trouble with the PSU, case clearance and mobo throughput if you plan on adding another 3-slot 3090 - those are all sub-1k though so you can always swap once you need to!

[tryonenow]

AWS GPU compute is pretty cheap if you don't need their newest beefiest rigs.

[bm1362]

I've prototype on a collab notebook and have been using AWS GPUs for training. Having to jump through hoops for quota and the general anxiety in trying to optimize the costs led me to just build a machine with a 1-time upfront cost. As a sole proprietor, I can write off the capex against future revenue anyways and depreciate it.