Cuda over VPN?

is it possible?
i do a lot of work over vpn to my office, however, as cuda is tied to the gfx drivers, it does not seem to want to work, as vpn invokes a softwear based gfx driver…

are there any workarounds?

By VPN do you mean Windows remote desktop? This won’t work, because as you say it uses a software graphics driver.

If you’re just doing console work you could use telnet or ssh. VNC also works:

I have run programs on the card via remote desktop. I mean that I run them on the card on the remote machine.

What “won’t work” exactly?

Philip, do you have 2 cards in your system? as i only have one… it could be that yours works using the secondary if its not attached to the desktop? or were you not really talking about remote desktop?.. VNC could be an option… but that would leave the machine open and visible at my workplace, which is not ideal amongst 200 people.

Simon, will the cuda drivers soon be independant of the gfx-drivers?

i imagine that will be the case, for your tesla and server products due soon,… any eta?


If it won’t work, do you have another suggestion? Besides VNC, Ultr@VNC …etc. (I can’t use it)

I need to compile and execute some CUDA files on my lab PC, but sometimes I wouldn’t like to go there.

Thank you :)

Why can’t you use VNC? Is it firewall restricition or something?

I’m using Remote Admistrator 3 by Famatech to perform remote operations on CUDA machine. It works fine and it requires much less bandwidth than VNC’s I’ve tried.

When I launched VNC and connected, I just saw a black screen and no other signal on screen.

OK, I will try to use Remote Admistrator 3.

Thank you.

I work using SSH, and you can also use remote desktop using X server, but that’s linux. I have problems with the examples that display the result on the screen, since my graphics card is not a G80.
It’s working for me :)

What is your graphics card?

Could you tell me?


Does this Remote Administrator work for you also?

FWIW at NVIDIA we use VNC and it works fine for console work. Graphics work is a whole 'nother ball of wax.

I work from home in the UK on a VPN and VNC into machines in California for testing all the time. Windows and Linux.


@Mark. video drivers are not loaded in the session space when we do a RDP to a remote machine(Win2k3) They are not loaded even if one does a console RDP. Is it the limitation of nvidia drivers or is it that there is some constraint from Win2k3 architecture?

It is a limitation of RDP itself. RDP uses own video driver and this is why you can set arbitrary screen resolution for RDP session, even if present hardware does not support it.

I wouldn’t even call this a limitation. RDP was designed with other things in mind.

Thanks AndrieB for the information…So now another thing which pops up in my head is, is it possible to do the same stuff on Linux system, i.e using the graphics hardware remotely and also how GPU then handles multi user case??..I guess RDP equivalent for linux might not be having MS RDP’s limitation?..

If you are running programs with no OpenGL-interop, then you can just connect to the Linux system with SSH. I do this all the time. You can even run multiple CUDA programs at the same time. I’m pretty sure different users at the same time would even work, but I haven’t tested that. The problem with multiple users is that all the context switching (only one kernel executes at a time) slows down both CUDA programs quite a bit. You really don’t want to share a GPU between two processes unless they are both intermittently using the card.