Can you use Amazon EC2 GPU instances for real-time rendering?

5
votes

I need a remote PC/server which has a decent 3D card in it, to perform real-time 3D rendering... imagine running a 3D game on a remote server and that's a good comparison.

Most VPS and dedicated servers do not have good graphics capabilities for obvious reasons but Amazon do have special GPU instances. They're sold for GPGPU computation, using the GPU for data-crunching using tools like CUDA, but I wondered if they could also be used for real-time 3D rendering.

Can anyone provide a solid answer to that?

Edit: I should add it's my own 3d code and I want to know the capabilities of EC2 for this purpose, not a generic EC2 question

2
cudaamazon-ec2gpgpu
Anyone running Windows EC2, as a side question what graphics does it report? – Mr. Boy
Perhaps @ArchaeaSoftware will comment. Amazon EC2 GPU instances have Tesla M2050 GPUs in them. The main focus of these is GPGPU compute tasks. It's generally possible to do linux graphics offline rendering (OGL) with these also, but the GPU isn't really configured for windows graphics rendering work. Even for windows GPGPU compute type activities, the best experience is had with Windows Server 2012 (or Win8). By the way, these types of questions, while interesting, aren't really well-suited for the SO question/answer format. You might have better responses on superuser or amazon forums. – Robert Crovella
If they are offering you cuda-capable GPUs in the server, then you can make real-time rendering. GPUs are for that. – eLRuLL

2 Answers

2
votes

Amazon GPU servers are equipped with NVidia Tesla GPUs .While these are best at doing GPGPU they also have more than average capabilities for real-time graphics rendering.Though in this respect they are inferior to NVidia GTX cards (see GPU specs on NVidia website).

Now , you can use Amazon for real-time rendering but your bottleneck will be the network bandwidth.Tesla cards can be used with OpenGL to render graphics into Offscreen buffers very very fast , but then you should find the way to read pixels for each rendered frame and stream it to the client with acceptable frame rate.OpenGL pixel read from GPU is already very slow (though you can do some hacks using PBOs ping pongs) but I don't really think you can stream pixel packages ,with standard resolutions (800x600 or even less)from remote server ,so that the client gets it at minimal acceptable frequency.I do believe it will be possible in the future :)

P.S My answer is based on personal experience with Amazon EC2

0
votes

Yes, amazon EC2 is well-suited for rendering. I've been doing this at large scale for over 3 years for a mobile application. Throughput has been fine for short animations which I move from EC2 to S3/CloudFront.