GPU Transcode - Multiple GPU's

[pir8radio 1292]

Does emby (or rather ffmpeg) support transcoding on a server with say, 4 nvidia GPU's in it?  Or will it only utilize one GPU?

[Luke 37022]

Hi, yes it does. The upcoming 3.6 release will have improved support for this.

@@softworkz

[softworkz 3326]

The upcoming version supports multi-GPU operations. 

 

That means that it will detect all available GPU's - each with all of its supported codecs.

It will allow you to select different accelerations for each codecs, even multiple in a priority order.

But the initial 3.6 release will not include load balancing capabilities in a way like "gpu1 is already performing 2 transcodings, so choose gpu2".

 

This has just been postponed due to time constraints, though. We've been working hard towards enabling this, and there's just some small steps missing.

Thus I'm hoping that we'll be able to deliver that functionality very soon after the 3.6 release.

 

 

Regarding Nvidia hardware acceleration, you need to be aware that they are limiting the maximum numbers of concurrent encoding and decoding sessions - especially for consumer boards.

That's why you might be better off using one or two of their professional boards instead of four GeForce boards.

 

Also you shouldn't start doing calculations to estimate transcoding performance based on their 3D or compute capabilities: NVENC and NVDEC are utilizing dedicated IC's on their boards that are independent of their 3D performance (with the exception of cuda/opencl hardware filters).

[pir8radio 1292]

The upcoming version supports multi-GPU operations. 

 

That means that it will detect all available GPU's - each with all of its supported codecs.

It will allow you to select different accelerations for each codecs, even multiple in a priority order.

But the initial 3.6 release will not include load balancing capabilities in a way like "gpu1 is already performing 2 transcodings, so choose gpu2".

 

This has just been postponed due to time constraints, though. We've been working hard towards enabling this, and there's just some small steps missing.

Thus I'm hoping that we'll be able to deliver that functionality very soon after the 3.6 release.

 

 

Regarding Nvidia hardware acceleration, you need to be aware that they are limiting the maximum numbers of concurrent encoding and decoding sessions - especially for consumer boards.

That's why you might be better off using one or two of their professional boards instead of four GeForce boards.

 

Also you shouldn't start doing calculations to estimate transcoding performance based on their 3D or compute capabilities: NVENC and NVDEC are utilizing dedicated IC's on their boards that are independent of their 3D performance (with the exception of cuda/opencl hardware filters).

 

I'm looking at a few quadro P4000 cards   their cross reference shows "unlimited" for concurrent sessions.   

[softworkz 3326]

I think that's a good choice.

 

Start with a single one to see how it performs.

After Emby 3.6 is released, you can add a second one and we can work together on the load balancing.