By now I am pretty much convinced that the problem has to do with:

1) Win10 itself, or
2) the WDDM overhead, or
3) the driver, or
4) a mix of above

In my mind, the possibilities are:

a) It's actually working correctly, in a situation where the code is moderately to heavily cpu-bound and there is no tangible benefit toward running at a faster gpu clock

or

b) Stupid NVIDIA driver is stupid.

....

It is very likely to be b.

Unfortunately, after the change from CUDA 65 to 80 a few days ago, and the compulsary driver update from 376.53 to 381.65, the problem has become even worse.
It looks like as if crunching above the default clock 1152MHz in no longer possible.

For example, I had recent tasks run at 1280MHz and around 77% TDP. After a few minutes, the clock fell back to 1152MHz, TDP around 55%.

Until a few months ago, it was no problem at all with this card to crunch at 1420MHz and about 95% TDP.

What's going wrong?

Still today, I will have to update driver on the PC with the GTX750ti inside, with which I have been experiencing the same problem for a while.
So I am afraid that also there, the problem will become even worse.

It would really be interesting how many other crunchers now have the same problem with a GTX970 in combination with the new software acemd_918.80 and the new driver 381.65.
Here, even mild overclocking makes the GPU jump back to default clock 1152MHz :-(((

Have you tried using a different tool to do your overclocking? I overclock with MSI Afterburner, and highly recommend it.

It would really be interesting how many other crunchers now have the same problem with a GTX970 in combination with the new software acemd_918.80 and the new driver 381.65.
Here, even mild overclocking makes the GPU jump back to default clock 1152MHz :-(((

I have one which is OC'd and doesn't display that behaviour.