CUDA 9.2 Win10 64bit older hardware (x79) with multi-GPUs seems unstable

I had been using Win 10 (ver 1709) with 9.1 mostly fine and stable. Then auto upgraded to 1803 by Microsoft (as it does). Around that time, thought nothing much of upgrading to CUDA 9.2. Then started having lots of instability (freezing intermittently, including mouse frozen, and after a minute or more, would unlock … to happen again at some point). TdrDelay etc. all set correctly.

I thought it was Windows … so tried full reinstall of Win10 1703 and then Win10 1709 … same problems. Realised it might be CUDA 9.2 … so went back to CUDA 9.1 (which also included reinstalling everything), but with Win10 1803 now (was 1709 originally). and back to stable again. So I think it’s CUDA 9.2 … and I think it’s the new packaged GeForce drivers, although I’m not sure. I usually don’t upgrade or change the GeForce drivers that come with the CUDA package. I also think it might be because of the older hardware (x79).

Update : apparently it is the latest GeForce drivers … they don’t configure Message Sending Interrupts and therefore any PLX motherboards are very unstable. Fix is not released yet but promised soon. Not sure if this is the actual fix, but apparently so …

Update : confirmed, it’s a lot more stable with MSI (Message Signalled Interrupts) mode deliberately enabled for the GPUs in Windows 10 and a PLX motherboard. You currently need a Windows utility to do this (can be found on the internet).