ACR/Lightroom GPU Acceleration

[Ron Scubadiver]

My desktop has an I5-2500 overclocked to 4.5 ghz and a Nvidia 750 GTX OC GPU.  Today I noticed ACR runs noticeably faster doing simple editing with GPU acceleration disabled.  After doing some looking around I found Adobe recommends at least a 760 (960 in current production).  These US$200 cards are almost twice as fast as the one I have, leading me to believe the difference between having a relatively fast CPU and a $200 GPU would be marginal.  A 970 GTX, which is a US$350 part would probably show an improvement over the CPU, but I am not sure how much.

Does anyone have experience with this issue or know of some timed benchmarks?

Later I am going to try stitching a pano to see what happens.

[zuglufttier]

A GPU should always be way faster since it's all about parallel tasks. If you do tasks that cannot be parallelized it doesn't matter what you use. In the end, your CPU will have a lot more power for those tasks.

I have the same CPU and a GTX770. I never changed settings in lightroom but with other editing software such as darktable, being able to use the GPU was a difference of making the program usable at all or not

So, if you want compare performance, you should find out which tasks are parallelized and which are not. Another suggestion might be that something is wrong with your graphics card, even an integrated onboard card should make things faster than only using the CPU.

[Ron Scubadiver]

After looking around I could not find a clear answer, but your comment about integrated graphics providing an advantage over CPU was seen somewhere else.  Also, it is more important for the card to be recent than fast.  This may be a perceptual thing.  GPU is slow on the initial loading of an image, but faster zooming in and out.  It is unlikely there is something wrong with the graphics card.  At this point I am pretty much convinced a new graphics card is not going to make much difference.

[afx]

A GPU should always be way faster since it's all about parallel tasks. If you do tasks that cannot be parallelized it doesn't matter what you use.

It is not just parallelization. There is overhead involved in getting stuff into the GPU. So the workload needs to fit to make it worthwhile.
In general, the GPU support in many imaging apps is overrated and often buggy. In AfterShot and Capture One it is best turned off, too many weird effects on the output.

cheers
afx

[Jørgen Ramskov]

In general, the GPU support in many imaging apps is overrated and often buggy. In AfterShot and Capture One it is best turned off, too many weird effects on the output.

That's odd. When I used C1, I had it enabled and didn't experience any issues with it, quite the opposite, the difference in speed was quite noticeable.

[zuglufttier]

Yes, a recent card is always a good thing in this case! But then, the difference between an integrated intel card or a dedicated nvidia card should not be that big. Of course, again, considering that both cards support the same subset of functions that are needed for a proper acceleration.

Maybe nowadays more powerful GPUs may have become more useful but that was not the case the last time I checked. Anyway, the GPU acceleration in lightroom is a very nice thing.

[Andy]

A GPU should always be way faster since it's all about parallel tasks.

As a general statement, this is unfortunately not correct. There are many factors impacting the cross-over point when a parallel system is faster than a serial system.

To name a few:
1) Amdahl's law.
2) The relative speed difference between the serial processor and the parallel processor. Ron has a very fast "serial CPU" and a relatively speaking "slow parallel GPU". This ratio shifts the cross-over point where the execution of the parallel part on the GPU shows a reduction in overall execution time further out.
3) The efficiency of the implementation.
4) The communication overhead between CPU and GPU. As the GPU does not have an embedded serial CPU, it needs to ask always the CPU for further work or needs to send data back to the CPU for further processing
plus many more factors are impacting this statement

In the end, minimizing the wall clock time counts. For example:
If Ron's system with his OC CPU would finish a task in 20 seconds(CPU only) and the addition of the GPU option with the GTX 750 would not change the result, it is still 20 seconds.
If someone else has a very slow Celeron CPU taking 100 second for this same task and add a super fast GPU like the GTX 980, the overall improvement could be 4x (as example) = 25 seconds. Great from a speedup perspective in itself, but it is still slower from an overall time perspective than the first option.

rgds, Andy

[Jørgen Ramskov]

Anyway, the GPU acceleration in lightroom is a very nice thing.

I haven't noticed any big improvement in LR since Adobe added support. I noticed a much bigger improvement when it was added to C1.

[afx]

Anyway, the GPU acceleration in lightroom is a very nice thing.

I haven't noticed any big improvement in LR since Adobe added support. I noticed a much bigger improvement when it was added to C1.

Lightroom still does not support the GPU for processing (OpenCL), so there is no acceleration in the processing.  They just finally use the GPU via OpenGL for the rendering like the rest of the world. The crop tool is still jumpy compared to others though.
C1 and AfterShot in contrast can use the GPU for processing via OpenCL.

cheers
afx