EC2 GPU Instance on Windows

Question

I am having a lot of trouble installing CUDA Toolkit on Amazon EC2 GPU Instances. I tried using the HVM Windows 2008 Server AMI from Amazon and also the NVIDIA GPU image both launched on g2.xlarge instances.

I connect to the instances through　RDP (remote desktop protocol).

I downlaoded CUDA 5.5 Toolkit for windows vista 7 8 DEsktop 64 bit from NVIDIA website. Everytime i try to install this toolkit I get the same error from NVIDIA installer: Couldn‘t find NVIDIA compatible card.

Is it because I connect through RDP ? I try to follow : http://adnanboz.wordpress.com/2012/01/06/how-to-set-up-amazon-ec2-windows-gpu-instance-for-nvidia-cuda-development/

But i'm stuck when installing CUDA Toolkit.

AMazon documentation is very sparse for windows :

http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/using_cluster_computing.html#install-nvidia-driver-windows

Again installing driver complains it cannot find NVIDIA card whether I launch the driver install program connected through TightVNC or through RDP.

Any pointers ?

Answer 1

I did not have any difficulty installing CUDA Toolkit 5.5 or 6.5.

• Launch a clean Windows 2012 R2 AMI on the g2.2xlarge instance. (The clean AMI also lets you use spot-priced G2 instances, unlike the NVIDIA CUDA AMI.)
• Tip: use a security group that permits both TCP and UDP connections on port 3389. This lets latest versions of RDP work faster.
• Install Visual Studio 2013 (I believe "Visual Studio Express 2013 for Windows Desktop" is ok for compiling 32-bit apps, but install a trial of the Pro edition to be safe).
• Install the CUDA Toolkit (do not install any driver except the one that comes with the toolkit, since this often causes problems)

To run applications:

• Install Team Viewer or another VNC-type app
• Connect
• Launch an admin command prompt (right-click Command Prompt icon, choose Run as admin)
• Run command query session to display the current user sessions. It will look like this:

  SESSIONNAME       USERNAME                 ID  STATE   TYPE        DEVICE
  services                                    0  Disc
  rdp-tcp#24        User                      2  Active
  console                                     3  Conn
  rdp-tcp                                 65536  Listen
  

  Look for the line that starts with "rdp" and has your username. Often the session ID will be 2.
• Run command tscon <session ID> /dest:console /password:<password of current user> in order to move your user session from the RDP virtual video card to the hardware video card.
• RDP will disconnect (if you still had it connected), but Team Viewer will stay connected (the resolution may change).
• Run your CUDA program, start debugging, etc.
• Once the program is running, you can actually re-connect via RDP for better performance.

Alternate way to run applications:

• Open admin command prompt and find our your session id (as above)
• Run tscon <session ID> /dest:console /password:<password of current user> & <command to start your program> e.g. tscon 2 /dest:console /password:1234 & myCudaApp.exe
• RDP will disconnect and your program will start
• Re-connect RDP

Tip: Use Open Hardware Monitor to keep tabs on GPU usage.