Yes it is a huge problem the Pascal, and i don't think it will change soon,
i am with my Maxwell and happy for the final results (next step: bios mod for OC, they are watercooled so it is ok, but i don't know if the hack will like it!)
yes i used No-HT to make test cause i had and still have a problem (more cosmetic than other) about the Nvidia manager
* with HT (so 44 threads) impossible to access to it (error message) but hack 100% stable and reliable,
and i tried with imac 16,1, 17,1, as macpro 5,1, 6,1 definitions, nothing,
BUT the results are not bad at the end in Luxmark (i am in top 5, sometimes 2nd in some LuxMark benchs, on their website)
so only cosmetic, when i want to upgrade the Nvidia driver or cuda one, i just boot in Non HT
* with non-HT: strangely the hack is 5°C more hot on same benchs, and no so interesting in term of pure perfs
>>> i join to you some results on CPU benchs only with Luxmark
OKRASIT, did a powerful hack with 2 CPU xeons,
but on CPU ONLY benchs, the results in 2 are not so far than one CPU as mine
PUGET website described why: when your computer start to move from one CPU to the other one (adding tasks) there is a time where the perfs are not good at all, i think that a mono CPU (in my goal) is much better
and as you said: much problematic to manage for install
but i repeat Okrasit as Metacollin did it so clearly possible
the 2696 v4 is a good buying (on Ebay): i had it for the price of an old v3
One point important for people who use or will use NVme and specially Samsung series as SM961
you need at least 15 to 20% free on it
on my 1TB, if i dont have 200GB free, the writing benchs on AJA are close of 300MB/s instead of 2000MB/s
and it was not a problem of t°C at all: cause watercooled
just problem of free capacity
PS: i will shoot everything and restart form the scratch cause i made too many tests, and now my install is crappy in some ways, so i will re install everything from the scratch
good luck