Article Thumbnail

Adobe Premiere Pro CC 2015.3 Pascal GPU Performance

Written on August 18, 2016 by Matt Bach
Share:
Table of Contents:
  1. Introduction
  2. Test Setup
  3. Exporting to 1080p - Single GPU
  4. Exporting to 1080p - Dual GPU
  5. Exporting to 4K - Single GPU
  6. Exporting to 4K - Dual GPU
  7. Render Previews - Single GPU
  8. Render Previews - Dual GPU
  9. Conclusion
  10. Related Hardware Articles
  11. Recommended Systems for Premiere Pro

Introduction

With new models of video cards being launched and updates to Premiere Pro being released on a constant basis, we spend a significant amount of time here at Puget Systems re-doing testing that is at times only a few months old. In fact, although our last Premiere Pro GPU article was published only about 2 months before this article, the release of Premiere Pro 2015.3 (10.4) as well as the new NVIDIA Titan X (Pascal version) video card meant that it is already time for us to re-run our benchmarks. Most software updates don't result in significant performance changes, but sometimes - especially in conjunction with other hardware or driver improvements - there can be a big enough change that we find the need to adjust what we recommend to our customers.

In this article we will be looking at the performance of Premiere Pro 2015.3 when exporting and generating previews with the latest Pascal video cards from NVIDIA. Specifically, we will be testing both single and dual GPU configurations of the GTX 1060 6GB, GTX 1070 8GB, GTX 1080 8GB, and the Titan X 12GB video cards. In addition, we will also be testing the previous generation GTX 980 Ti 6GB to see if there is any significant performance advantages to using the latest cards.

Test Setup

For our test system, we used the following hardware:

Premiere Pro can heavily utilize the CPU, although in our CPU-based testing we have found that there is only a relatively small benefit to using a workstation with dual Xeon CPUs. Because of this, we are simply going to test with a single configuration using an Intel Core i7 6950X CPU. While there are faster CPU configurations you could use depending on if you export to 1080p or 4K+ resolutions, this is still a great all-around CPU for Premiere Pro which makes it ideal for comparing different video cards.

The different video cards we will be testing are:

To help with consistency - and since the benchmarks we performed ran for several days - we programmed a custom script using AutoIt to start Premiere Pro, load the relevant project, time how long it takes to export the timeline with the appropriate settings or generate previews, close Premiere Pro to clear any data from the system RAM, then loop while changing the project file.

The files we will be testing with came from a variety of sources:

1080P H.264/CineForm
4K H.264/CineForm
Provided by: Jerry Berg
Barnacules Nerdgasm - YouTube
ProRes 4K Grant Petty
Blackmagic Design Forum (available for public download)
4K RED RAW Provided by: Mike Pecci
Director & Photographer
6K RED RAW Neumann Films
RED Dragon Test Shot (available for public download)

In order to make our testing as accurate as possible, we used relatively simple timelines for our testing. In the past, we've loaded on the accelerated effects to show the maximum difference between cards, but we found that this was not representative of real-world performance gains. Our test timelines consisted of:

  • 4-5 clips arranged in series to make a 60 second timeline
  • A basic transition (cross dissolve) was applied between each clip
  • Lumetri color correction effect applied to each clip
  • Vector-based logo graphic added to the bottom corner of the footage

Exporting to 1080p - Single GPU

While more and more people are starting to shoot in 4K and higher resolutions, 1080p is still by far the most common resolution to export to. Because of this, we thought we would start our testing by looking at how Premiere Pro is able to utilize a single GPU when exporting from a variety of resolutions and codecs to H.264 1080p:

Premiere Pro Pascal GPU Performance

Right off the bat, we are getting some very interesting results. In some instances (H.264 1080p, CineForm 1080p, and RED 4K) we saw essentially no difference between any of the cards we tested. Others (ProRes 4K) showed some minor differences, but they were only minor and not consistent  - the more powerful cards were sometimes slower than the less powerful cards.

However, there were three times where you would notice the difference with a more powerful video card. If you work with H.264 or CineForm 4K footage and export to 1080p, there is a definitely an increase in performance by using a GTX 1070, and a further increase by using a GTX 1080. While these two tests did not show a benefit to using a Titan X, Exporting RED 6K footage to 1080p did show a pretty significant decrease in export times when using a Titan X.

Exporting to 1080p - Dual GPU

Premiere Pro Pascal Dual Video Card
With two video cards, our results are quite a bit different than what we saw with a single GPU. This time, the only time we saw an advantage to using more power cards was when exporting RED 6K footage to H.264 1080p. Even then, even the GTX 1070 was able to match the GTX 1080 or Titan X.

What is really interesting here is that if you compare these results to the results from our single GPU test in the previous section, it becomes clear that we are definitely hitting some sort of bottleneck that is preventing the workstation from exporting any faster. In most cases, a single GTX 1080 is going to be faster than any dual GPU configuration (with the exception of RED 6K where you would need a Titan X to match the performance of two GTX 1070/1080 video cards). We are confident that we are not hitting a storage bottleneck (we tested with faster storage to make sure), but it is possible we are hitting a CPU bottleneck even though we are using one of the fastest CPUs currently available for Premiere Pro. No matter what is causing us to hit this performance wall, however, our testing indicates that with current hardware there is rarely a need for a dual GPU configuration if you are primarily concerned about performance when exporting to 1080p.

Exporting to 4K - Single GPU

Although 4K isn't as widespread as 1080p quite yet, more and more Premiere Pro users are either exporting to 4K or are considering exporting to 4K in the near future. Exporting to higher resolutions is more taxing on the system as a whole, so it will be interesting to see if the difference between each video card changes when exporting to 4K versus exporting to 1080p:

Premiere Pro Pascal 4K performance

With the export resolution turned up to 4K, we actually saw no difference between the different video card models when exporting any of the 4K footage (regardless of codec) to H.264 4K. However, we did see a big difference when we scaled the RED 6K footage down to 4K. What is interesting is that most of the performance gains we saw when exporting to 1080p in the previous two sections was when we started with higher resolution footage and downscaled to 1080p. While not conclusive, this does suggest that a more powerful video card is more likely to be beneficial if you export to a lower resolution than your source footage.

Exporting to 4K - Dual GPU


With two video cards, there isn't really much to discuss. Exporting RED 6K footage to H.264 4K was the only time we saw a difference in performance, but even then there was no reason to go to anything more powerful.

However, if you again compare the single GPU results to the results in this section, you will notice that using two cards is rarely faster than using a single card. In fact, the only time it is faster is if you are using RED 6K footage. Even then, a single Titan X is as fast as any dual GPU configuration - even two Titan X's. Technically, two GTX 1070 cards would be cheaper than a single Titan X so there may be a cost reason to go with dual cards if you work with 6K+ footage, but at that resolution the additional VRAM on the Titan X is likely to be very useful. In addition, single GPU configurations tend to be more stable in the long term with fewer driver or software bugs popping up at random times.

Render Previews - Single GPU

Premiere Pro Preview Generation Pascal Titan X

Rendering previews is very interesting because - for whatever reason - the results are almost the opposite of what we saw when exporting to 1080p. This time, with the exception of ProRes 4K, the different video cards were all able to generate previews for the 4K and 6K footage at approximately the same speed. In fact, only the 1080p footage (both H.264 and CineForm) showed a benefit to using a higher end GPU.

Render Previews - Dual GPU


If you read the earlier sections, this is probably starting to look familiar. In a couple of instances there is a pretty big benefit to upgrading from two GTX 1060s to two GTX 1070s, but no benefit to using a GTX 1080 or Titan X. However, if you once again compare these results to the single GPU results in the previous section, it is clear that for generating previews there is rarely a need to use a dual GPU configuration. Two GTX 1070 are certainly cheaper than a single Titan X and gives nearly identical performance when working with 1080p footage, but the downsides to multi GPU configurations are very likely to outweigh the difference in cost for most users.

Conclusion

If there is one thing we learned in our testing, it is that the performance difference between the cards can change drastically depending on what source codec and resolution you use as well as what resolution you are exporting to. So if you tend to only work with one of the codecs we specifically tested, we highly recommend simply looking at those results and ignore everything else.

With that said, however, there are a few broad conclusions we can make based on our results:

  1. In most situations, there is no need to use a dual GPU configuration. In some situations, dual GPUs can be faster than a single card, but we found that a single higher-end card (typically a Titan X) was able to match any dual GPU configuration we tested. If you happen to work with the type of footage where we saw a benefit to multiple video cards and are on a tight budget, you may consider using two GTX 1070's over a single Titan X. However, at that price point you are likely to find the additional VRAM on the Titan X to be very useful. In addition, single GPU configurations tend to be more stable in the long term with fewer driver or software bugs popping up at random times. 
  2. When exporting, a faster GPU is more likely to give you a performance benefit if your source footage is a higher resolution than the export resolution. We saw a very distinct trend where the different model of video cards performed identically when the source and export resolution was the same. This doesn't mean you will always see a benefit from a faster GPU when you export to a lower resolution than your source footage, but in four of the six benchmarks we ran where this was the case, there was a very distinct performance advantage to using a higher-end video card.
  3. Generating previews saw the biggest variance between the different GPU models with 1080p footage. ProRes 4K also saw a benefit, but all the others (H.264 4K, CineForm 4K, RED 4K, and RED 6K) saw virtually no difference in the time it took to generate previews with the different video cards.

Keep in mind that the more GPU accelerated effects you use, the larger the difference between each of the models we tested should become - so if you tend to use a lot of the accelerated effects the GTX 1080 and Titan X should provide an even larger benefit than what we showed in this article.

Related Hardware Articles

If you are configuring a workstation for Premiere Pro, we have a number of articles regarding the hardware requirements for Premiere Pro that you may be interested in. If you prefer a summary, we also have a list of Hardware Recommendations for Premiere Pro based on the findings of these articles.

Premiere Pro CC 2015.4
Storage Optimization

How many drives should you use and how should you configure them?

NAB Webinar: Choosing the right hardware for your Premiere Pro Workstation
Comprehensive guide for what hardware to use in your workstation

Premiere Pro CC 2015.3
CPU Comparison

What CPU should you choose for your Premiere Pro workstation?

Premiere Pro CC 2015
Multi Core Performance

Does having more CPU cores give you more performance?

Premiere Pro CC 2015.3
Pascal GPU Performance

Will a more expensive video card improve performance in Premiere?

View All Articles

Includes older articles that may not be relevant for the latest hardware and software revisions 

Recommended Systems for Premiere Pro

 

Compact
Workstation

Purchase

Utilizing an Intel Core i7 CPU with up to 10 cores, this workstation provides the best possible performance for Premiere Pro in a compact package.

Standard
Workstation

Purchase

While physically larger than the Compact Workstation, this system allows for up to 512GB of RAM and a wider range of storage options.

 

Tags: GTX, 1060, 1070, 1080, Titan X, Premiere, Performance
Fabio Palma

Fantastic article. Comparison to the previous article, am I right than now there is no issue anymore with GTX1080 vs Red4K footage? There are tricks to handle Red 4K footage with GTX1080 ?

Posted on 2016-08-24 05:44:20
Djslee

How is mac pro comparing to windows machines. Is there a big differt with the same cpu. Can you compare that???

Posted on 2016-09-08 09:44:37

That is actually something we've wanted to test for a while, but haven't had a good opportunity to do so. At some point, we'll probably just have to bite the bullet and buy the highest-end Mac Pro to test with.

I will say that I expect a Mac to be much slower per dollar and have a lower top end of performance. Macs are still using the Xeon V2 CPUs (I believe) which are two generation old now (which usually means about a 20-25% performance difference per clock), and the Radeon cards they use should be quite a bit slower and hotter than the newer Pascal cards we used in this article. So hardware-wise, I would expect a $8,000 Mac Pro to be slower than a $2,500 PC (not to mention what you could get out of a $8,000 PC!). The big question is how much the OS comes into play and that can't really be answered without directly comparing a Mac Pro to a PC workstation.

To keep things fair, we might wait until the next Mac Pro is released. The current one is pretty dated hardware-wise so while it would still be relevant to test with the current model, it really would be better for us to wait for a more fair comparison I think.

Posted on 2016-09-08 17:48:11
Dan

Really useful article! Same question as a previous comment that doesn't seem to have been answered. I'd be very interested in seeing how the RX 470 and RX 480 compare to the GTX cards, especially a 1060, with OpenCL vs CUDA etc.

Posted on 2016-09-13 11:50:36
Scott Crozier

It would be interesting to see this test done with a LUT / Lumetri applied to each clip. Basic encoding/decoding is done on the CPU for most codecs (except RED gpu debayer etc) in Premiere so that largely explains your level results. Scaling is done on the GPU as well so that's why the scaling tests are more evident. You probably aren't going to see a big difference until you apply some effects. Lumetri is probably the most used effect so it would be an interesting test. Also if you ran a program like GPU-z on the side as well I'd be curious to know what the GPU usage is in these scenarios. I know you probably don't have time for this but I think this test would be more enlightening if there were some effects applied.

Posted on 2016-09-20 23:42:14

I believe we did that. Here is the description of the testing we did, quoted from up in the article:

"4-5 clips arranged in series to make a 60 second timeline
A basic transition (cross dissolve) was applied between each clip
Lumetri color correction effect applied to each clip
Vector-based logo graphic added to the bottom corner of the footage"

Posted on 2016-09-20 23:46:30
Scott Crozier

Missed that list somehow. Weird. Would have expected a bit more ooomf out of the Titan X. I have a 1080 so I guess no real complaints out of me.

Posted on 2016-09-20 23:48:56

Edit: William beat me to the reply by a couple minutes, sorry for the double info!

Each of these tests actually did have a Lumetri Color Correction applied as well as a cross dissolve between the clips. This is mentioned in the Test Setup section, but looking back at the article, that isn't super obvious. Maybe I should have something on the charts as well pointing this out. Either way, rest assured that we did do our testing with some accelerated effects! We even considered trying to show more of a "maximum difference" in performance by loading up the clips with a ton of other accelerated effects, but we felt that doing so wouldn't be a very realistic benchmark for how people actually use Premiere Pro so we stuck with just color correction and basic transitions.

Logging the GPU usage is a great idea - I'll keep that in mind for next time. It might not mean much to our readers that aren't quite as hardware-savvy (there are actually a surprising number even for these benchmarking articles!), but that is something I could put in a collapsed section or something like that for those that are interested in those fine details. Thanks for the suggestion!

Posted on 2016-09-20 23:53:05
Scott Crozier

No problem! Thanks for doing these tests! Some very good info here. I'm considering upgrading my 4770k system to a 6xxx system and your articles on this site have been very beneficial!

Posted on 2016-09-20 23:58:01
Pal Pacher

Great and useful article as always. I am using PuSy workstation: dual Xeon E5-2680v3 2.5GHz with dual Quadro K5200, Win 10. I was wondering if replacing the Quadros with a single Pascal Titan X would significantly improve the performance of the system in Premier Pro CC using HD and 4K? thanks

Posted on 2016-09-23 05:47:01
david smith

Interesting article - but it would have been interesting to see a CPU only render time on each graph - to show the advantage of GPU rendering in each scenario.

Posted on 2016-09-28 09:59:12

Hey David, there are a couple reasons why we typically don't test CPU-only performance on things that can utilize the GPU as well. The first is that we have limited time, and testing CPU-only would just about double our testing time since CPU-only should be somewhere around 6-10 times slower for the timelines we use. That wouldn't be too bad in and of itself, but we already have a huge laundry list of testing we want to do - both for Premiere and other applications - and knowing how much slower not using a GPU would be simply isn't very high on that list.

At our core, Puget Systems is a workstation manufacturer so our articles have to prioritize testing the hardware that we would actually sell to our customers. The performance gain you get in Premiere Pro by using a GPU is so large that we would pretty much never sell a Premiere Pro workstation without one. So while it might be interesting to know how slow CPU-only would be, it really isn't all that directly useful for us. Even though we prioritize testing that we feel is most important to ensuring we are selling the right hardware to our customers, we've always felt that our testing is extremely useful to the community at large - which is why we publish them freely rather than keeping the results to ourselves. Unfortunately, that does mean some topics (like CPU-only performance) we likely won't be testing any time soon.

Posted on 2016-09-28 18:31:08
Igor

Is it possible to do a test with different disks setup configuration in the test hardware, considering to the current one disk which is used for the OS project file, Media File, Media Cache, Export ... I think it's a bottleneck. Therefore, all results are equal, with very little variations. The test system does not allow the full potential of graphics card. That's why we get the similar time value with 6k exports in 4K and 1080p from same card. And of course much greater mutual disproportion, especially at higher resolutions. I think I would had a more reasonable results with at least 3 disk configuration: 1 OS, 2-Project Files and Media Cache, 3-Media Files.

Posted on 2016-10-06 10:32:15

I'm actually right in the middle of doing a bunch of testing on storage options for Premiere Pro. I'm only about half way done with the testing, and it is going to take another week or so to finish. In addition to exporting and rendering previews, I decided to add importing, conforming audio, generating peak files, and testing how many simultaneous video streams can be played back live so it is taking about a full 8 hour day to complete a round of testing for each storage configuration.

One thing I can tell you from the data I have already, however, is that for what we tested in this article (exporting and rendering previews) a single SSD is more than fast enough to not be a significant bottleneck. These actions are simply CPU/GPU limited at the moment so storage speed is not a significant factor. The only time I'm seeing lower performance is if you have the projects, media, cache, and scratch all on a platter drive - and that only results in about a 10-15% hit to performance. Multiple SATA SSDs with the projects, media, cache, and scratch spread around and even NVMe drives (which are about 5x faster than SATA SSDs for read and 3x faster for write) show no difference compared to having everything on a single SATA SSD.

The times when spreading out the different files across multiple drives helps the most looks to be when importing, conforming audio, and generating peak files. I'm still not done with all the testing, but at the moment it looks like the ideal storage configuration is a primary drive with just the OS and Premiere Pro installation, a fast secondary drive for your projects and source media, and a fast tertiary drive for your media cache and scratch. I'm not sure if you would see a benefit with a NVMe or RAID for those second and third drives quite yet but it looks like there might be a small benefit. Again, this will only help with Importing, conforming audio, and generating peak files. For exporting and rendering it doesn't appear to really matter as long as your projects, media, cache, and scratch are on some sort of SSD (even all just on your primary drive)

Posted on 2016-10-06 19:05:25
Igor

I agree that single SSD is more than fast enough to not be a significant bottleneck. But only for HD. For 4K RED RAW and 6K RED RAW I think is botlleneck for sure. And those 1080 and Titan X cards are "screaming" for that kind of task. Anyway, great article! Enormous work and effort as always. Two thumbs up!

Posted on 2016-10-07 08:40:11
Gregory Bickford

Wow. I can't tell you what a great service you're providing to indie filmmakers who build their own systems. Thank you so much. About the Pascal GPUs... The 1080 and Titan X seem pretty close, but with the advent of 8K material (I'm soon to be an Epic-W owner) how much more effective will the Titan X be? Will it's 12GB vram help more with Premiere previews, and put it clearly above the 1080? And what if we have more than one video layer in the timeline? Would there then be a significant advantage to the TitanX? Thanks again!

Posted on 2016-10-21 05:07:12
Dave

Some excellent information on here, maybe I could get some advice. If I were to use three ssd's and two hdd's in my workstation how would I go about it to set this up for best results when editing with adobe premiere pro cc? I have two regular ssd's and one 750 series pcie ssd, the latter I would like to use as working media disk. The 1st ssd as bootdrive, the 2nd as a scratchdisk. Then I have two sata hdd's that I would like to use as a final export or save disk to be used in raid-1 (so mirrored) I am a complete noob is there something I should know? Motherboard used is an Asus X99 Deluxe II with a single Xeon E5 1650v4. The fast 750 series intel pcie ssd is 1.2tb and the bootdrive samsung ssd is 512Gb, the scratchdisk is a samsung ssd of 1tb. I will be using windows 10 home or pro edition and I have 64Gb of ECC Reg RAM. Any help on the best structure would be highly appreciated. Graphics card used will be the GTX 1080.

Posted on 2016-11-18 00:29:58

Hey Dave, storage configuration is something we get asked fairly regularly and I'm actually working on an article right now covering storage configurations in detail. Overall, your setup sounds pretty good except I would move the media cache onto your scratch drive as well. Just to cover everything, this is what I would generally recommend doing with a setup like yours:

OS & Software: Either the 512GB SSD or the Intel 750 1.2TB. Using the Intel 750 will make booting Windows and launching programs a bit faster but likely nothing mind-blowing

Projects & Source Media: Either the 1TB SSD or the Intel 750 1.2TB. The Intel 750 would let you import media ~20% faster than an SSD and is about 50% faster for generating peak files. Personally, I would use the Intel 750 here rather than using it as your OS drive, but nether way is really wrong - it just depends on where you want/need the difference in performance.

Media Cache & Scratch: Use the remaining SSD for the media cache and scratch files, but especially for the media cache. This makes a massive difference for importing (up to 6x faster import times). Having the scratch separate from your projects doesn't actually do anything for performance, but having it on this drive helps even things out a bit and makes it pretty easy to clear your media cache and scratch files whenever they start to get too big.

Export & Archive: Platter is fine here. Even a fairly slow platter drive (~125MB/s) isn't going to make more than a percent or two of difference when exporting compared to using an SSD. No need for RAID froma performance standpoint, but a RAID-1 (mirror) like you mentioned can be good for hardware redundancy. Just remember that RAID is not a backup solution! It protects against a single disk failure but does nothing if the controller fails, the RAID configuration gets corrupted, you get a virus/malware, or if you simply accidentally delete something.

Hopefully that helps you out. It sounds like you already had a pretty good idea of what would work well, but I definitely would move the media cache files off the primary drive (which is the default) if you can.

Posted on 2016-11-21 19:59:29
Dave

Thank you Matt, for you're elaborate answer. To be fair I not only don't have any experience building a pc nor do I have any experience shooting or editing video with adobe premiere, I just decided that's what I wanna do so I've got a pretty steep learning curve ahead:) after seeing the process of some of the computers been build at Puget Systems I am trying to emulate some of it into my first ever workstation. Since I cannot order a Puget Systems pc from Europe. (bummer)
I went out on a limb with choosing the ECC variant ram modules after reading on you're website (Puget Systems) that ECC RAM is unofficially supported on Asus X99 Deluxe II boards (combined with a Xeon processor) Is there something I need to do to make that work (in the bios settings) or will it just work? I have asked numerous people about it even Asus themselfs wouldn't or couldn't answer me if it would work, they can only go by the book. Anyway I decided to go for it. It's pretty exciting and a little scary since everything is so bloody expensive. After reading you're advice I changed my mind about the two big hdd's (2x 4Tb WD black) in raid-1, if the safety of loss is so low I better use them as two separate drives at least I then have a full 8Tb of extra memory. I think the idea is to export anything worthwhile outside of you're system anyway wright?

Posted on 2016-11-22 12:54:30
Richard

Hi,

Thanks for this informative article.

I am looking to buy a new workstation right now.

Currently, I am thinking of the following setup:
- 1 dedicated GTX 1080 for rendering previews and accelerated effects (i.e. no monitor attached to card for display purposes)
- 1 extra GPU (built-in on CPU or separate) for display of GUI Adobe Video Apps on a 4k computer monitor
- 1 Blackmagic Design DeckLink Mini Monitor 4K for 10-bit full screen output of video to a separate 4k color calibrated reference monitor

Here are my questions:

1 - Will my system perform worse if you choose to connect 1 (or 2) 4k computer monitors to the GTX 1080 compared to using that card for computational power only?
2 - If I go for a dedicated GTX 1080 for computational purposes can I rely on a built-in GPU on an i7 CPU for my 4k display monitor or should I buy a separate video card and go for the non-GPU built-in CPUs like the 6900K?
3 - if a second GPU, should this be identical to the other card? On the Adobe hardware forum, I read that PPro only maximizes the weakest GPU available. Is that true or false?

Please advise.

Posted on 2016-11-28 20:35:45