CUDA over Remote Desktop now available for Tesla GPUs

John Melonakos Announcements, CUDA 5 Comments

Update: Jacket over Remote Desktop is now available for Quadro devices too! Read this post. Jacket over Remote Connections is also documented extensively on the AccelerEyes Wiki.

Over the past several years, many Jacket programmers have requested support for Remote Desktop in Windows.  We are pleased to report that recent NVIDIA drivers now enable Jacket to run over Remote Desktop, for some system configurations.

nvsmi

Enable TCC Mode for NVIDIA GPUs

Specifically, the requirements to make this work include:

  • Windows Vista, Windows 7, Windows HPC Server 2008, or Windows HPC Server 2008 R2
  • The latest NVIDIA driver (as required by Jacket)
  • Tesla GPU
  • TCC-mode enabled on at least one (Tesla) GPU
    • To enable TCC, the Tesla cannot be connected to a display. This means you need to connect the display to another GPU.
    • The GPU needs to be an NVIDIA Tesla device.

TCC is currently enabled using the Windows command-line function NVSMI, which is installed with the NVIDIA driver.  The image to the right shows how to enable TCC-mode for your GPUs (click the thumbnail to maximize).

Finally, please make sure your jlicense.dat file includes the text TS_OK (terminal services are OK).  If you are missing the TS_OK feature in your license file, simply download a new jlicense.dat from the Manage Licenses page.

Other remote protocols, such as VNC and SSH, continue to work well for remote Jacket operation.

Comments 5

  1. It is worth noting here, that all Tesla GPUs are not TCC-enabled by default. This means all GPUs are not ready by default to run Jacket over Remote Desktop. A rule of thumb is, if your Tesla GPU does not have a video out, it is TCC-enabled (example: the 1060 series).

  2. Readers may want to note that the non-TCC devices are not displayed on machines when accessed through a Remote Desktop Session. For example, here's the output on my machine with a GT8400 and a Tesla C1060:

    >> ginfo
    Jacket v1.7 (build 4ba10c6) by AccelerEyes
    License Type: Designated Computer (on 64-bit Windows)
    Licensed Addons: DLA, SLA
    Multi-GPU: Licensed for 4 GPUs

    Detected CUDA-capable GPUs:
    CUDA driver 263.06, CUDA toolkit 3.2
    GPU0 Tesla C1060, 1296 MHz, 4096 MB VRAM, Compute 1.3 (single,double)

    GPU Memory Usage: 4058 MB free (4096 MB total)

  3. If you have multiple cards, use the '-i' parameter to select a specific card.  On my machine with a Quadro2000 and Tesla2050, I initially got this error:
    C:Program FilesNVIDIA CorporationNVSMI>nvidia-smi.exe -dm 1

    Unable to determine display mode for GPU 0000:0F:00.0: Not Supported

    Specifying the second device (zero-indexing) got it working:
    C:Program FilesNVIDIA CorporationNVSMI>nvidia-smi.exe -i 1 -dm 1
    Set driver model to TCC for GPU 0000:42:00.0.
    Reboot required.

Leave a Reply

Your email address will not be published. Required fields are marked *