Nvidia driver update causes console to hang/crash when CUDA support enabled

[K1FSY]

Hello! I was having an issue with the cuda fft support where it seemed like the cuda driver would fail and the cpu driver would take over. This was noticeable because the waterfall noise floor would move faster in cpu mode, and the noise floor increased about 2dB. Stop/Start of the radio would always switch it back to cuda and drop the noise floor. I was noticing some other lag/latency today so I checked dpc checker and sure enough nvidia kernel driver was popping up with some long waits. I upgraded the driver to the latest 546.33 game ready driver. The release notes for a driver version between 536 and 546 specifically mention fixes for DPC latency on ampere cards. This driver has completely fixed the latency issues with the nvidia kernel driver, however Console will hang and crash if cuda support is enabled and works fine with it disabled. It won't even connect to the radio as it seems as though it does the graphics initialization first. I'm reporting this not so much because I'm having this problem but in case something has changed with the nvidia driver and needs to be worked around. Graphics card is an RTX-3090, CPU an AMD 5900X and 32GB ram. A "clean install" of the nvidia drivers were performed (and other 3d applications run fine.) Running most recent December beta of console.

[Simon G4ELI]

Hi,

Is this with the latest 3.3 beta? Not heard of this before, what radio and bandwidth are you using, I wonder if I'm exhausting available CUDA memory.

[K1FSY]

Hi Simon,

Yes this was was with the latest 3.3 beta. I'm using 384khz bandwidth with an HL2. It was working fine (as in it was actually working in CUDA mode) just prior to updating the drivers from and rebooting. Once I rebooted it hung when connecting the radio and I quickly tracked it back to having CUDA enabled, and can reproduce at will by enabling cuda which hangs and crashes the app.

[Max]

I upgraded the driver to the latest 546.33 game ready driver. The release notes for a driver version between 536 and 546 specifically mention fixes for DPC latency on ampere cards. This driver has completely fixed the latency issues with the nvidia kernel driver, however Console will hang and crash if cuda support is enabled and works fine with it disabled. It won't even connect to the radio as it seems as though it does the graphics initialization first.

As you are running HL2 and as I have not updated Nvidia drivers for a while I updated to 546.33 Game Ready. Actually I usually use the Studio drivers as I understand they are not so "edgy" and I'm not a gamer. I think the Game Ready drivers is another name for Beta driver in Nvidia terminology?

Did the update and strangely I too could initially not start HL2 immediately. Console seemed to freeze (not responding message). But after two or three attempts it now starts fine and continues to do so. TBH I think it's probably just a coincidence but just mentioning it.

Anyway I've now stopped and started radio and also other SDRs all with CUDA enabled and also tried again after rebooting the PC and all is well. No issues at all with or without CUDA. My proc is i9-12900K with Nvidia RTX3070.

Max