Haven't tried 30B models but love my P40s. Have eight of them in a single rig. They're nowhere near as fast as my 3090s, but I didn't pay much for them either.

Ik_llama.cpp is your best friend with the P40s. Being datacenter cards, they support P2P, which ik enables and uses during inference. This speeds things up considerably vs vanilla. In my experience, 2x faster or even more if talking about prompt processing.

Cooling them is much easier than most think. First, they're not power hungry. They'll happily idle at 8-9W each. Second, you can power limit them to 170-180W without a noticeable degradation in performance. Third, and my favorite, the PCB is the same as the 1080Ti FE/reference or Titan XP. So, you can slap any waterblocks for these two cards onto the P40. You'll need to cut a 1.5x1.5cm piece out of the acrylic/POM for the EPS connector. Alternatively, you can desolder the EPS connector and solder two 8 pin PCIe power connectors. I went the dremel route because it cuts the cabling in half.

Do yourself a favor and get an older Broadwell Xeon with DDR4 or even DDR3 motherboard to go with them. Those Xeons come with 40 lanes and are dirt cheap, as are their motherboards. While I don't see more than 6GB/s per card in my rig running minimax at Q4, I'm sure I'd get an additional 1-2t/s if I had an x16 connection to each GPU.

I'm sure many have seen this half a dozen times already, but here's my P40 rig for the umpteenth time: