Take the 2-minute tour ×
Video Production Stack Exchange is a question and answer site for engineers, producers, editors, and enthusiasts spanning the fields of video, and media creation. It's 100% free, no registration required.

While looking into buying a new graphics card for the editing, live playback and encoding of videos in the new 64bit Premiere Pro CC, I was wondering about:

1) how and when a graphics card's memory is utilised or loaded?
2) when would one benefit from the higher graphics card Memory?
Would I benefit during editing/live playback, or when encoding and exporting a video?

As of summer 2014 the Nvidia GTX 770 seems a sweetspot in terms of price/performance that grasped my attention. This model is abvailable in 2GB and 4GB versions. I will be using some filters and will be using Magic Lantern RAW 1920x840 @ 24 fps footage timeline for short videos of around 4 minutes. Thanks!

share|improve this question

3 Answers 3

1) Everytime the GPU has to process something, it will load the necessary data (in that case your video frames) into the VRAM. Simple as that, the GPU cannot work with your system memory. Though the GPU gets it's data from the system memory, so your system memory will act as a buffer and you don't need to load that much data at once into the VRAM.

Though programs which offer a real-time workflow will want to load as much data into the VRAM as possible. That's the reason why you can allocate a specific amount of VRAM in Adobe After Effects (not 100% sure if you can do the same in Premiere). So every time you apply an effect in Premiere that is GPU accelerated, the GPU will have to request data into the VRAM to process the frames.

New versions of OpenCL and CUDA are improving the way of how the GPU will get its data and reducing the overhead caused by the CPU which will reduce the amount of VRAM needed at a time but VRAM will always be a very important factor with the current PC architecture. Current GPUs use GDDR5 which is A LOT faster than DDR3 system memory. This is needed as the memory speed is a very important performance factor in GPU computing. Overclocking your VRAM can give you quite a performance boost in memory intensive 3D rendering applications (though probably not in Premiere). Current GDDR5 based GPUs like the GTX 780 Ti are capable of more than 300 GB/s (336 GB/s in this case). DDR3-2133 is substantially slower with a peak bandwidth of 16,66 GB/s per module. Also 2133 is rather uncommon in OEM PCs, DDR3-1600 and 1300 is more common.

2) So you will benefit from a large amount of VRAM everytime you work with very high resolution images or many videos at once aswell as with high-frame rate footage. Premieres rendering engine is utilizing the GPU for both the preview aswell as the final rendering if you chose so.

In practice I'd say it's nonsense to have more than 2GB of VRAM for Premiere unless you work with other applications at the same time that utilize the GPU. Processing bitmaps doesn't require all that much data at once unless you work with crazy resolutions like 8k or 16k, the processing capability of your GPU will be the bottleneck before your VRAM.

Thats my opinion based on what I know about the Premiere video engine and GP-GPU. Premiere might do some other stuff in the background that requires more VRAM. So if you have a very specific use case I'd rather make a benchmark or search for one before making a purchase decision. If you just want to make a general guess this information should help you quite well to make a decision.

share|improve this answer
    
It is worth mentioning that it does depend on the architecture being used. There have been numerous graphics cards over the years that worked closely with system ram and even a fair number that directly used system ram as their storage space. Fundamentally, the proximity of the memory just involves how quickly the GPU can get data. It can either request data be sent from RAM to itself or it can work directly from the local GPU memory, the only difference is that it doesn't have to talk to anything else to get access to GPU memory. –  AJ Henderson May 28 at 20:03
    
Ah, it does appear that your statement is true for NVidia GPUs though that will be changing with the Maxwell chips. –  AJ Henderson May 28 at 20:07
    
Well yes thats what I wrote. At the end the data still has to be in the VRAM though, even if its for a short amount of time. Even the new CUDA versions aren't doing anything new in that regard, they just improve the way system and GPU memory is handled. It's very important that the active workload is in the VRAM as its substantially faster than the system RAM. Maybe I should add that to the answer, thats an important factor to know. –  Professor Sparkles May 28 at 23:02

I have read that the extra ram (4 gb vs 2 gb) might be utilized when you are working with 4k footage, but there are a lot of other factors at play here (cpu / data throughput etc). This is a good article about choosing the right components for your editing pc (including graphics card): http://ppbm7.com/index.php/tweakers-page/83-balanced-systems/94-balanced-systems

share|improve this answer

I haven't seen any difference between 2GB and 4GB. My GTX680 2GB is practically the same as a friend's GTX680 4GB in terms of PP performance.

I don't know if this is just on our systems, but even with a fast GPU PP just can't handle more than a couple of filters without dropping frames.

It's kind of infuriating, really. My GPU can create beautiful landscapes with hundreds of filters from nothing in realtime but something simple like a few layers of text is too much for PP to handle.

Hint: try DaVinci Resolve Lite. It's free and benefits immensely from a good GPU!

share|improve this answer
    
Note that your GPU doesn't fully render those landscapes and there aren't hundreds of filters. Games pre-render most of the stuff and use lots of tricks to make it so that shaders operate at high efficiency rather than high accuracy. They also operate on polygon models rather than raster data. When using a GPU for video acceleration, it has to do all the calculations from scratch, without the advantage of pre-calculated optimizations, on raw raster data. It is a far, far harder problem. –  AJ Henderson May 26 at 21:21
    
@AJHenderson Sure, PP rendering isn't at all similar to game graphics, but it could be way better. While PP starts dropping frames with only a few color correction filters, Resolve renders practically the same thing in 3x realtime. –  Pichan May 27 at 7:46
    
What kind of footage are you working with? Even on my system that doesn't have GPU acceleration, I can handle several color correction filters on 1080p footage in real time without dropped frames in PP. –  AJ Henderson May 27 at 13:19
    
@AJHenderson Usually h264 straight from the camera. I know it isn't the optimal format but again, Resolve handles it effortlessly. I may have exaggerated the problem a bit, but it really doesn't take much to make PP slow down. –  Pichan May 27 at 14:08

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.