hopefully someone will do full benchmarks of the rx480 vs nvidia cards. most of the benchmarks i have seen are limited to just a few scenarios that usually don't give the full picture. some older benchmarks showed amd cards performing about 50% worse than nvidia cards, but some newer benchmarks are only around 20% slower. the rx 480 might be closer to the gtx 980 in games, but it might be somewhat slower than the gtx 970 in adobe apps if that handicap still holds true. its price tag of $200 is pretty good, but nvidia may price the gtx 900 series cards to compete. also, while the list of cuda only software is shrinking, its still something to be aware of for anyone considering amd cards.
you are also comparing the gtx 1070 against a slower rx 480, when the gtx 1060 comes out it may be closer to the rx 480. you need to figure out how much performance you need or want vs your budget.
Many articles show, including a Linus TechTips Video, that there is almost no performance difference from a r9 290, 970, 980, 980 ti, fury, r9 380 fury x etc. Anything from mid range up is pretty much the same, but like you said many benchmarks don't really show the whole picture, like Linus's is exports only.
The performance difference I'm seeing from last gen are within 1% of each other and I don't see it changing that much this generation. I hope to be proven wrong, but right now it seems a $229 RX 480 could be the better buy for video editing.
any test showing little to no difference between gpu's is bottle-necking on the cpu or other hardware.
the chroma keyed cineform export showed some difference in the hardware. it shows the r9 390x performing in the middle between a gtx 970 and 960. the r9 390x in games performs closer to the gtx 980, so that suggests an amd handicap, at least in that scenario. the rx 480 had an early game benchmark released, showing it just slightly below the r9 390x.
If you watched the video I linked you would know the test bench setup is hardly bottlenecked in any fashion.
i have seen that video before. there is always a bottleneck in the sense of a weakest link, not as in a slow system. premiere pro is primarily a cpu bound program. its not until gpu based effects are added to the timeline/clips and/or certain codecs like cineform are used, that gpu usage starts to become a factor.
It's not a bottlenecked system. Sorry. You're just saying that UT not actually providing any proof. If the test system was bottlenecked we wouldn't see any change from card to card, which we do. Also the gpu results would be similar to cpu only, if it was bottlenecked, which it is not.
as i said before "any test showing little to no difference between gpu's is bottle-necking on the cpu or other hardware". that means one component or another is limiting performance in that one test or scenario. linus even speaks to this starting at 4:30 in the video, how the gpu usage was low. so now you are asking for proof? its in that chart at the same time, those gpu's are very different in power to have such similar times, meaning there is something else holding back the software from running faster (bottleneck), ie the cpu. continuing in the linus video at around 4:50, he goes on to say, the third test finally showed them what they were looking for. as in it was a test that finally used enough gpu power to make a difference and get away from the cpu bottleneck, and he actually mentions "cpu encoding". my bottleneck statement wasn't geared towards the last test, which i acknowledged earlier was using the gpu's more and showed a difference. i used that last test's chart to speculate the rx 480 performance against other cards in my previous post. if you still do not understand what i was trying to say, that's cool, we will have to agree to disagree.
I'm pretty sure Premiere Pro doesn't really scale well past 8cores and I'm pretty positive Linus used a mutli-core xeon. You could be 100% right but i was under the impression he had plenty of HW power for that test. Why present something as fact, like he did in his conclusions, if the benchmarks are bottle necked and the data really doesn't show anything?
It seems like a giant waste of time to produce a full video and go through all the trouble to present bottle necked numbers.
Again, it just doesn't make sense to me.
the problem is there are many many different software components to premiere. so in a worse case scenario it might use 2 cores max, in another it might be well over 12. so around 8 or possibly 10 is kind of the happy medium. so far benchmarks on the 10 core i7 are showing it working very well in premiere over the 8 core, but still hard for many to justify that price tag. its the same with the gpu benchmarks or test scenarios, sometimes gpu acceleration is hardly being used so different gpu's make little to no difference and sometimes premiere needs alot of gpu power for another scenario. that's why i wanted to see more well rounded, or full, gpu benchmarks with amd vs nvidia, to really see in which scenario's amd cards perform different.
he did have plenty of hardware power for some testing to be useful, but he did very limited testing. you have to also remember, linus covers all sorts of tech, so he doesn't specialize in any one thing. kinda like the saying, "good at everything, great at nothing". his staff members edit his videos, so he probably barely knows the basics of premiere and just ran some tests until one (the third and last one in the video) finally did something he could report on. the first two tests could be seen as a waste, as they are bottlenecked results. but between the first two and last test results it does show that there are difference scenarios where the gpu wont be used as much, and any descent card will work fine.
his conclusion statements might be taken in different ways, he goes from saying there isn't enough difference to recommend faster cards and then saying that if you are using different tools (software/plugins) it might be different (as in a faster card might be better). thats a big if that he doesn't fully explain and for many people its the reason why they need fast and sometimes even multiple gpu's with premiere and/or other software they might be using on the same computer. he concludes on a statement of diminishing returns, which could be taken as diminishing returns in premiere, or for prices of the gpu cards like the gtx 980 and above...
Hey thanks for the great insight and info sir. I really appreciate reading posts from people now know what they're talking about. Lots to think about. Probably going to wait and see for benchmarks.
Also, by no means I am disagreeing with you or saying you're wrong, I'm just trying to understand what you're saying better. Previously you said the system was bottlenecked which gave us the similiar results across the board. And then in your last post you acknowledge he had "plenty of power". So how can he have plenty of power and still be bottlenecked?
as i have previously tried to clarify, i first said: "any test showing little to no difference between gpu's is bottle-necking on the cpu or other hardware", and in the very same post went on to say the third test showed some difference. so i was suggesting only the first two tests were bottlenecked results, and the third was not bottlenecked.
as for the "plenty of power" statement, i said "he did have plenty of hardware power for some testing to be useful, but he did very limited testing". so now you are trying to combine parts of two separate statements, one of which i've tried to clarify at length in a previous post. he did have "plenty of power" to do testing, but not plenty of power for the first two tests. again, linus even said the gpu usage was around 30% for the second test. therefore something was limiting/ holding back/ bottle-necking/ preventing the gpu usage in the first two tests from being higher. so he would have needed a faster computer to allow premiere to put more demand on those gpu's in the first two tests or to simply use premiere in a way that would use the gpu more, which they did in the last test.
while the third test may suggest to some the different cards don't perform much differently, that is only a visual trick of the chart, as the "cpu only" result lumps all the gpu times together. looking at the 980ti vs the 960, the 960 time is almost 80% longer. that is a major difference not clearly represented on that chart in a visual way.
anyone looking at the rx 480 might want to google "AMD RX 480 pcie slot power"
there are some reports from review sites that its drawing too much power from the pcie slot on the motherboard and several reports from users with fried pcie slots and dead motherboards. it seems to early to know how big of a problem this will be for the reference card, but hopefully won't be a problem for partner custom cards with 8 pin power connectors.
in other news, the gtx 1060 is to be released soon and supposedly targeting the rx 480. looking at the cuda cores i suspect it will perform somewhere close to the middle between the gtx 970 and 980. the rx 480 seems to be matching the gtx 970 and r9 390 in game benchmarks.
The Puget benchmark is very informative, but I didn't reached their same conclusions. They usually get impressed by 8 to 13% speed improvements. I don't think these kind of gains are noticeable, even when rendering long movies.
So don't forget they sell systems, and draw your own conclusions before getting any new GPU or even CPU for your workstation.
Not the same software, but some guys here are comparing the Blackmagic Resolve (candle benchmark test). Their conclusion was that the Polaris it the right choice, at least for Resolve.
Considering Premiere Mercury engine likely uses the same kind of features of the GPU than Resolve, probably Polaris will be a good choice for Premiere too!
A little late to the party but maybe someone in the future will find this helpful or semi interesting. I had the same question as the OP and couldn't find anything definitive elsewhere and was finally able to do some tests myself. These are not really scientific, but are real world for what I do (wedding and corporate video) and they answered my questions while creating some new ones.
HP Z820 Windows 10 pro
Dual Xeon E5-2665 8-core @ 2.4Ghz/3.1 boost
64Gb RAM running @1600 Mhz
Samsung 850 256Gb for OS measured at 500+ Mbs read/write
Samsung 850 1Tb for Media Cache measured at 500+ Mbs read/write
4Tb Raid 0 for project files measured at 550+ M
The test: Rendering a 10min time line created in the latest Premiere CC (2015.3 I think) and sent to Media Encoder. Time line is about a 50/50 mix of 4k and 1080p form a Sony A7s II, timeline is 1080p with lumetri and Filmconvert on every clip (both GPU accelerated. the first 10 seconds is a dynamic link to an AE project that is very CPU dependent. The render is to a highly compressed 720p H.274 clip for web delivery. GPU is monitored by GPU-Z.
Card Time to Render Max GPU Load Max Memory Used
GTX 970 FE 6:43 87% 1505Mb
GTX 1070 FE 6:32 82% 1941Mb
RX 480 (x1) 6:41 100% 6124Mb
RX 480 (x2) 5:00 100%/100% 5207Mb/4621Mb
Durring all renders, CPU maxed out between 50-60%, RAM used maxed between 55-65%
To me it looks like there's a bottleneck somewhere or I didn't enough GPU accelerated effects, but if that was the case, wouldn't the dual RX 480's perform the same as the others? I also find it interesting that more memory was used on the AMD's than Nvidia's.
This didn't solve the problem for me. I like the render times for the dual 480's but was concerned that they would only scrub a 4k timeline at the same speed as the GTX 970 I'm replacing since only one GPU is used for real time playback. The issue is when I edit I stack multiple camera angles on the timeline and was struggling with playing more than one at a time. The one 480 will play the 3 layers I need, I didn't try more. The GTX 1070 will too.
Thinking about dropping Adobe and moving over to resolve/fusion so I did the candle test others have linked to in this thread. I didn't get the same results but this is what I got:
Blur Nodes: 9 18 30 66 Max @ 24fps
GTX 970 FE 21 12 7+ 3+ 8
GTX 1070 FE 24 18 11+ 5+ 12
RX 480 (x1) 24 12+ 10 5 9
RX 480 (x2) 24 13+ 11 5+ 11
This is on Resolve 12.5 Lite, the the dual RX 480 is using one for the GUI and one to render, in studio with both rendering I think they would out perform the 1070.
Either setup will be great for me and the cost is about the same. No question the GTX 1070 is a more powerful single card, will run cooler, and use less energy, But I feel my system is using the RX 480's more effectively unless GPU-Z isn't reporting the stats accurately.
Not sure what I'll do yet so I'll have the cards for a few more days if anyone has a better way to benchmark them I'm open to it, I also welcome any feedback/advice, I often lurk on the forums but never post.
As a side note, both drivers have options to push 10-bit color.
thanks for sharing your findings. the rx 480 seems to be holding its own with premiere pro. with the rx 480 memory usage it suggests the 8gb version should be used with premiere. the different memory usage probably has to do with cuda vs opencl and premiere's history of not being fully optimized for amd/opencl. the most impressive part of the premiere times to me is the dual rx 480 showing the biggest change, while the gtx 1070 shows only a minor change. dual gpu might be performing faster as each gpu can work on separate frames and allow for quicker turn around for the cpu's to continue working.
depending on the project and media, premiere can max out cpu cores as early as 6 cores (12 threads). so having 16 cores (32 threads) may be causing the lower cpu usage between 50-60%. you could try disabling hyper-threading to see if performance increases in premiere. davinci resolve should have a better chance of maxing out the all dual xeon cores/threads.
several benchmarks show anything around the gtx 1060 and rx 480 to be fine for most projects, until working with red 6k. so you could test 6k r3d with lumetri and a few random gpu fx, to see if the playback performance and export times show bigger differences between the video cards. RED Dragon Test Shot (6K Downloadable file) - YouTube
you seem to have a lot of experience. It's a bit off the topic but it regards the gtx 1060.
Before (using Premiere Pro CS6)
CPU I7 5820K at 4 GHz, RAM 16GB and GTX 760. (SSDs plus 2 raid 0 HDD)
Mostly doing DV, but also multiple camera Full HD. No problems.
DV to mpeg2 DVD CPU maxed out.
DV to H.264 or HD to H.264 CPU between 40 to 70 %
Just recently 4K that's when things start to slow down, but still 2 4K videos on timeline (each one hour) Second video is masked out with garbage mate 8 or 16 point and fastblur to overlay parts of the second video into the main video. Still can scrub or play at fast speed (2x) CPU after a few minutes or so maxed out. When I encoded it to H.264 I noticed the following:
CPU maxed out, RAM maxed out and GPU maxed out. (Have All CPU meter and GPU monitor to observe)
Then I decided to upgrade RAM to 32 GB and GTX 760 replaced with Asus GTX 1060 6 GB turbo.
Do a similar project, but this time I have to move the 16 point garbage mate overtime. Just to adjust it over time the system gets quite unresponsive. CPU constantly maxed out RAM about 70%. In the beginning of the timeline it was easy to adjust, towards the end after 40 min it was very difficult. Normally you can use the mouse to position the 16 points of the mask, but this time, I had to key in the values, so that I could see the moves.
I am writing this, for you mentioned in your above post, that the gtx 1060 should be fine for most projects.
It encodes then fine 1 hour timeline takes approximately 2 and a half hour (Now after the upgrade the CPU lies around 40 to 50 %, GPu around 30% RAM 70 to 80%)
So the question is, Do you have an idea where there could be the bottleneck? It doesn't seem to me to much of a burden.
thanks for your time and ideas
How did you manage to make RX 480 work in rendering? I've recnetly bought the card and I am frustrated to say the least... Media Encoder, Premiere and After Effects don't use GPU acceleration on this card no matter what settings I've used... If anyone can help with this issue I'll be forever grateful for that.
If you would like to test with my Premiere Pro BenchMark (PPBM) you are welcome to download it and test your CS6 system. I am no long posting the results but will attempt to get back to you personally. There are four tests, one is CPU intensive, one is Disk I/O intensive and the other two are GPU accelerated.
2 people found this helpful
ephtrackx, The first thing I would check is the mercury playback engine (project settings) and make sure its on GPU - opencl, then I'd make a short time line in premiere with any footage you have available and put about 10 gaussian blurs on it set to 100. With your GPU monitor open (I've been using GPU-Z) play back the time line and watch the gpu usage. This test puts minimal load on the other aspects of your system so it's unlikely to bottleneck while asking a lot of the gpu. If you still don't have any activity I would completely remove the drivers and reinstall them. Use 16.7.3 and avoid 16.9 until it's no longer a beta. I have had instances where premiere wouldn't release the graphics card even after being shutdown, but I at least had activity on the GPU rather than none. In another scenario, renders with media encoder where significantly slower if I left Premiere open in the background. I'm sure you've all ready covered most of that, I can't think of much else that would cause it not to use the GPU.
RoninEdits You are right about the hypertheading, I do get a small rendering boost when it's off, but only with the GTX, not the RX480, and only with projects that have a dynamic link to AE.
thanmartin look for a setting in preferences that says "use Intel fast h.264 encoding" or something like that, I think it's under general, maybe memory settings. I turned it off and it made a big difference for me.
I did some further bench marking, not so much real world, but to find the actual differences in performance. Same system as above, hyerthreading off just for fun.
One minute timeline of 4k XAVC from a sony A7s II. 10 Gaussian blurs set to 100. Render to 4k h.264.
GTX 970- 5:12, full resolution playback of 3 active
GTX 1070- 3:28, full resolution playback of 6 active
RX480x1- 4:27, full resolution payback of 4 active
RX480x2- 2:15, full resolution playback of 4 active
CPU and RAM stayed between 10-20% and all the cards used about 90% of their available RAM (8gb RX 480's). I've been doing tests like these for the last week and they all have similar results, whether I used DNx HD vs. XAVC, or rendering to 720 or 1080.
What I get out of it:
The Cuda vs. Opencl in Adobe argument isn't as valid as it once was. The RX480 is suppose to perform slightly better than a GTX 970, and it does. I think if Adobe still heavily favored Cuda, the RX480 would perform closer to or worse than the 970 given their specs. I could be wrong so feel free to correct me.
I also see Adobe making good use of dual GPU's for rendering with the pair of 480's nearly halving the render times of a single card. Having seen other benchmarks and stats I would make an educated guess that the pair would equal to or beat a single GTX 1080.
What does it mean in the real world? Nothing. I found that on a real project the difference between all of the cards is negligible. I can add Lumetri, Film convert, do some scaling, add anything else thats accelerated and it might take the GPU to 20-40% when playing back a full resolution. Drop that to 1/2, and I can playback the clip with 10 gaussian blurs on it just fine with both the GTX 1070 and RX480. Rendering out a 13 minute video for a client is 30 seconds faster with the dual RX480's taking 10:41 compaired to the GTX 1070 at 11:16. I also like to stack clips in the time line when I edit and I found that I'll max out the CPU decoding the the multiple streams of XAVC before I come close to maxing the GPU. If I switch over to Resolve, the 1070 does better but only because I don't have Studio and can't use more than one GPU for rendering, Still the RX480 is usable for the same reasons it's usable in premiere. Maybe if I had a newer faster system things would be different.
So in deciding between dual RX 480's or a GTX 1070, I lean toward the 1070. Cost is about the same but the Nvidia drivers seem to play better with Adobe, rendering is the only time the second GPU is useful (for me), and only if I have several effects that can take advantage. The 1070 also uses less energy and runs cooler. I like the idea of multiple graphics cards because I think we'll start seeing more software that can take advantage of it. By the time it's common place though, the next, next gen cards will be out and I'll be upgrading again anyway.
Just sharing my experience with first world problems, hope it helpful and I welcome any feedback.
thanks for sharing some more testing results, looks like the rx 480 is performing very well in premiere. this is a good sign for amd's vega to compete with higher end nvidia cards.
dannonnicks, thank you so much for your response. I've already figured out, that I was misled by thinking that GPU accelerates everything, even just transcoding to another format without any effects (that what I was trying to do before). Apparently, that isn't the truth, and now I can clearly see GPU being used in previews and rendering with the gaussian blur test you've suggested. Sorry for being such a noob, just recently bought a decent GPU to try and experiment with GPU acceleration and do not know all the small things...