For preview the calculations may be being done to what is known as "single precision" while for rendering it uses a more precise "double precision" for better quality.
Single precision calculations would allow it to use the single precision CUDA cores on the graphics card while double precision would go to the double-precision cores. There are 32 times as many single precision cores in the "consumer" cards:
On the GP104 an SM combines 128 single-precision ALUs, 4 double-precision ALUs providing a 32:1 ratio
(From wikipedia)
Gamer or "Consumer" graphics cards tend to eschew double precision cores in favour of single precision cores because they offer the most "bang for buck" and are more than suitable for most home gaming or graphics tasks.
Workstation class cards such as the Tesla are more geared towards scientific applications and would have a full compliment of double precision capable units.
From Wikipedia:
The theoretical double-precision processing power of a Pascal GPU is ... 1/32 (of single precision performance) on GP102 and GP104.
The 104 is a GTX 1070, but I would expect a similar penalty to double precision calculations on the GP106 based 1060 cards.
Your CPU on the other hand probably does not have the same problem and single or double precision does not have the same performance penalty that the graphics card has.