my goal was to find out will it work like real GRID K2 so i can share gpu to multiple vm-s.
without mod with qemu-git on latest debian it works flawless...
Using qemu-git with what, exactly? KVM? Xen?
Are you saying that you can pass the GTX680 to a VM without modifying it onto a Quadro/Tesla/Grid?
with gird k2 ID, i can run it with grid k2 drivers,even with desktop ones, but restart, shutdown, POST screen doesn't work... but when gpu works it works same like 680gtx, in games, streaming hw encode... etc... with both drivers with id GRID K2...
Again, with what hypervisor? What do you mean by "restart, shutdown, POST screen doesn't work"?
If you mean that you don't see the POST screen in the VM on the GPU output, then this is normal. You will only get the SeaBIOS POST screen if the card is passed through as the primary (VGA) card. This means that QEMU has to dump the BIOS out of the card and side-load it for SeaBIOS to initialize. I have never seen this work on any hypervisor.
Restarting, shutting down and starting up again works just fine for me using Xen, with no GPU crashes.
Desktop and Quadro/Tesla/Grid drivers are actually the same if the version is the same. Quadro/Tesla/Grid drivers get tested much more thoroughly, hence why they are released much less frequently, but the binaries should be the same between those and the desktop drivers of the same version.
so pretty much card works with GRID K2 id... so next i wanted to try to share it...
There's should be a "pretty much" in there. I've used this setup extensively with Xen with no unexpected issues at all. If you mod to a Grid K2, your video outputs will be limited to SL-DVI modes. If you read a few pages back, this seems to be a known firmware/driver bug. If you want to fix it you will need to take your chances and flash the Grid K2 firmware onto your 680 (if your 680 is 4GB you shouldn't need to edit the firmware for RAM size, as discussed a few further pages back), but IIRC the K2 has no physical video outputs, so you will lose the use of those, and you'll only be able to use the card for GPU sharing with video output going out to a remote client.
on xen-server 6.2 latest - with any GRID K2 host driver... it detects it normally, but when i try to start vm with shared gpu... i got 3 lines in dmesg like...
its init and binding gpu, taking ownership and then nothing... xen-center crashes with error iomem... vgpu exited unexpectedly...
passing it like single card works sometimes...
I think it would be particularly useful to be able to compare the results against what happens with a real Grid K2. Xen GPU sharing is still sufficiently experimental that you could be dealing with a software bug or hardware compatibility issue. Can you confirm you use a different GPU as the dom0 primary?
on native windows 2012 R2 hyper-v installing with any GRID host driver crashes host (kernel i guess, mouse shows here and then with bsod sometimes, on keyboard numlock blinks)... all 3 outputs have some problems like black/white screen blinking and black/green snow (lol)...
Again, this could be related to the known Grid K2 firmware/driver bug.
There is no way installing the driver should be crashing the host unless there is a failrly epic software bug in play.
last epic try:
qemu with win 2012r2 vm with grid passed... i can make driver and hyper-v to work but i don't have gpu listed in hyper-v options (maybe cause of passtrough, so it doesn't detect it like phy gpu)
I'm pretty sure nobody ever tested nested virtualization like this, nor do I think anyone expects it to even remotely work.
so next step would be updating bios? would that help? is it possible to recover bios if it fails?
You could try. Worst case, you should always be able to re-flash the BIOS by attaching SPI lines directly to the EEPROM chip. I believe there is a discussion 20 or so pages back here about doing it using a Raspberry Pi. But you should only need to do that if the BIOS cross-flash fails. Because Grid K2 is a multi-GPU card, it will come with 2 BIOSes. You should probably use the primary BIOS from the K2 and you will need to edit the bus device arrangement. I think I wrote a post about where this is in the BIOS many months ago on this thread. Otherwise you won't be able to flash the BIOS onto the card.
did anyone managed to get it to work like real grid? is it even possible?
I haven't even heard of anyone who managed to get a real Grid K2 to work with anything but ESX, and even that is mostly from marketing material hearsay. Incidentally, ESX let's you do GPU sharing with a Grid, Tesla, and even some Quadros (Fermi 5000/6000).