Add specific flags for RTX 5060 and document the changes

[laurentdavid]

By looking at the open source version of nvidia driver 610 (and running the project through AI actually), I found several additional flags that together make it work on a dual RTX 5060 Ti.
Very experimental at this stages but it passes the cuda-sample tests simpleIPC and simpleP2P

[benbenbenbenbenben]

FYI I'm working with a pair of 5080s and the 5090 fix wasn't sufficient there. Your fix for the 5060 Ti did work however. Thanks for this.

One curiosity however. I had to reboot without a monitor attached to get it to work, but once the OS had started I could connect multiple displays and it would still work! Maybe there is a critical moment at startup where we can tweak something such that it allows the GPUs to peer correctly then resume business as usual?

[laurentdavid]

Hi @benbenbenbenbenben,

Thanks for your message. I must admit I did not try to reconnect the display after.

My guess was that the two cards have to be in the same setup/memory so allocation layout when we detect the P2P, so memory could be allocated from the same starting point. I guess you are right, but I am not sure this can be done without tweaking the driver itself. I am only beginning to understand the process, so I might be wrong. Great news that's working with other cards too !

Laurent