RTX 4090 and 4090Ti showcased and benchmarked :)

[bruce]

If the 4090 would perform twice or thrice a 2080 Ti, I would buy them.
Only if they can be used to the fullest extend.
At this time, very few (if any) WUs fully load a 3090, 3080, or a 2080 Ti even...
I think for folding and deep learning the low end 4000 series of GPUs will be excellent, just because I expect them to perform higher, while using even less power!
At this performance levels, the costprice isn't the biggest concern.
It's how many you can run, before tripping a breaker?

Science dictates how many atoms need to be simulated. There's a loose relationship between atom count and shader count. GPUs with too many shaders become inefficient because the fahcore can't increase the parallelism of small proteins.

[toTOW]

But there are plans to work around this, like running two (or more) trajectories at once on a big GPU ...

[Neil-B]

... and who knows - researchers are clever people - as the availability/capability of gpus that can run models with far more atoms increases there is a fair chance that projects will be created that actually push these ... for the RTX30 series there are now a number that fully utilise these - assignment challenges still remain but with a bit of luck (and a lot of hard work behind the scenes) that may get improved.