TNR massive slow in 12.5.6

Hi -

TNR in 12.5.5 was all realtime (also with "better" and "2 frames" at 100%). We never have to cache for that. But now after update to 12.5.6 (clean install) the TNR is massive slow down to max 8fps (1080p) - no realtime anymore. We have 4 GTX TITAN for compute, all driver up to date.

Candle Test:

09 Blur Nodes - 25fps
18 Blur Nodes - 25fps
30 Blur Nodes - 25fps
66 Blur Nodes - 25fps

01 TNR Node - 18fps (befor in 12.5.5 = 25fps)
02 TNR Nodes - 13fps (befor in 12.5.5 = 25fps)
04 TNR Nodes - 6fps (befor in 12.5.5 = 25fps)
06 TNR Nodes - 2fps (befor in 12.5.5 = 18fps)

Some one with similar issue?

Thanks in advance!

UPDATE

After this massive TNR slowdown we done a complete clean install with OSX 10.12.6 and DR14.1.1. But the TNR is still very slow! NVIDIA WEB, CUDA and BM DESKTOP driver are up to date and work fine, so no idea why DR have that issue with TNR only.

Nobody with the same issue?

CUDA 8 or 9? I think we're supposed to stay on 8...

We are on CUDA 8.x (Nvidia rec for SIERRA).

Strange: we work also on a WIN10 system with one TITAN only, but with more or less the same performance as our MAC system with 4xTITAN.