So with Pop Piasa's post here... http://www.gpugrid.net/forum_thread.php?id=5217&nowrap=true#56930

Here's a bit of promise for owners of Ampere GPUs: https://www.acellera.com/index.php/2021/04/06/release-of-acemd-3-4/

This new version brings support to the latest NVIDIA GPUs, including the Ampere architecture, as well as performance improvements. The simulation speed has been benchmarked against several systems at typical production conditions on different GPU devices (including GTX 1080, GTX1080 Ti, RTX 2080 Ti and RTX 3090). For the DHFR benchmark, on RTX 3090, ACEMD achieves the speed of ~1.3 µs/day.

This appears to be an explanation of why this project has not upgraded so far. Acellera is the owner and developer of the software, so we had to wait on them, not the GPUGRID team so far. Hopefully the license here will not need to be upgraded ($$$).

It looks like my RTX 2060 would be useful too, though I have never paid much attention to the Tensor cores.
https://towardsdatascience.com/rtx-2060-vs-gtx-1080ti-in-deep-learning-gpu-benchmarks-cheapest-rtx-vs-most-expensive-gtx-card-cd47cd9931d2

It would probably be more effective here than on Folding. I can use my GTX 1070's there.

It looks like my RTX 2060 would be useful too, though I have never paid much attention to the Tensor cores.
https://towardsdatascience.com/rtx-2060-vs-gtx-1080ti-in-deep-learning-gpu-benchmarks-cheapest-rtx-vs-most-expensive-gtx-card-cd47cd9931d2

It would probably be more effective here than on Folding. I can use my GTX 1070's there.

There is an advantage in not being on the bleeding edge. As I recall, it is not the first time it has happened here.
It just gets more expensive.

Also, while we are at it, I wonder how much work will be available? The AI technique has higher output, and may allow more of it to be done in-house.