For the last few years, I've been running >120k gpus to mine ETH. Before we turned them off after ETH switched to PoS, I tried really hard to find an alternative workload for them.
A large number of the GPUs were based on the PS5 chip, which is an APU (CPU+GPU) and ran as blade computers. It is actually a pretty power efficient chip, but the blades themselves are little use for anything that requires a lot of networking or data (they are diskless and ipxe boot a custom ubuntu).
I found a professor who's been doing research on various planetary data crunching efforts and we got his software (I think based on BOINC) running on these blades.
His software is analyzing data from telescopes to look for binary quasars. We crunched the numbers from our proof of concept and estimated that if we dedicated all the GPUs I have of that class, to this one task, that a 1 year of computation would be done in 1 month. At least two people have received Nobel's based on similar findings.
The problem was finding funding to pay for the operating expenses, megawatts of power, so it never got off the ground and the hardware has been turned off and decommissioned. Such a bummer.