Tim,
My email to you was not to belittle your efforts in any way, but just to put more illumination on the thermal processes. Putting an RPi, even with a fan, in a hot closet may not solve the problem.
>I've used lots of pi's and I've had them throttle because of thermal issues.
Both the Pis and Linux have had a lot of improvement over the years to handle thermal issues. For example, when you only have one core, turning off that core to let the SoC cool down has some issues. :-)
Most SoCs also have some thermal issues because the CPU and the GPU tend to be on the same SoC and therefore generate heat between the two. In a server system that has reduced or no graphics needs the heat sink can carry a lot more of the CPU's heat to the outside air. If you are considering doing AI work, rather than trying to use the GPU for that you could consider an accelerator (like we have discussed before) to spread that to another board.
>plus that's not good for an SOC in any case
A lot of this depends on the engineering of the SOC. There are a lot of issues other than just thermal that limit CPU clock speed and therefore limit the heat generated by the CPU. I cannot imagine a CPU that was not designed to run at maximum clock speed for all the cores from a thermal perspective, and of course turning it off makes it cool down again. If you tried to go into overclocking, you might have issues there, but then you have other issues too.
From what I have read the RPi definitely has some applications as a "desktop replacement", particularly in some of the areas where I operate. Solar panels and batteries could easily supply the power for running an RPi 5 with a reasonably sized LCD panel, important in various areas of the Amazon. In Uruguay there was one remote village where the five "One Laptop Per Child" (OLPC) units produced the only electric light at night, charged by solar panels in the day. My Lenovo W510 laptop (with a 130 Watt power supply) would have drained their batteries in an hour at most.
Another thing to consider is using the RPi as a "thin client", with the main computation is done remotely in "the cloud" and the Rpi is just used for graphical rendering and almost real-time capture of data for deep processing remotely.
Have a good time with your experiments. Most importantly, have fun.
md