AMD vs NVIDIA for Final Cut Pro X

[Asgorath]

I didn't mean horrible OpenCL in general, I meant horrible OpenCL that FCP uses. There is a difference. And the Titan X uses more memory and better speeds to overcome that maybe. I am sure NVIDIA beats AMD in some OpenCL tasks. But it looks like FCP is not one of them .

Then make sure you're a little more explicit when making complaints, since it sounded like you were saying OpenCL was horrible across the board on NVIDIA, which I'm suggesting is not the case. Show me a single-GPU score that beats the 14 seconds I've measured with my TITAN X. The point I keep making is that there are plenty of NVIDIA configs that run FCP extremely well. Maybe those are the 6GB+ configs, I don't know, but again I'll push back on claims that FCP runs horribly on NVIDIA across the board.

If all you're going to use your Mac for is to run FCP, then sure, sell your GTX 980 so you don't have to deal with the lack of boot screens and all that.

 

[koyoot]

My 12GB TITAN X runs BruceX in 14 seconds, per the other FCP performance thread. That's better than any single GPU result I've seen posted (edit: the 2 D700s in the nMP get a similar score). It also disproves your claim that NVIDIA is terrible at FCP or OpenCL. Here's another hint: if you compare 2 GPUs where one has massively more theoretical horsepower and get similar results, then the bottleneck is not related to raw GPU horsepower. This is performance analysis 101.

Fellow user on this forum gets BRuceX test in 8.9 on Dual D500 setup. In one of the BruceX benchmark threads it is.

 

[Asgorath]

Fellow user on this forum gets BRuceX test in 8.9 on Dual D500 setup. In one of the BruceX benchmark threads it is.

Did you miss the part where I was asking for single-GPU results? If I had 2 TITAN Xs, I'm assuming I'd get better than 14 seconds.

 

[koyoot]

Did you miss the part where I was asking for single-GPU results? If I had 2 TITAN Xs, I'm assuming I'd get better than 14 seconds.

I am not contradicting your statement

Just added something .

 

[Ethosik]

Then make sure you're a little more explicit when making complaints, since it sounded like you were saying OpenCL was horrible across the board on NVIDIA, which I'm suggesting is not the case. Show me a single-GPU score that beats the 14 seconds I've measured with my TITAN X. The point I keep making is that there are plenty of NVIDIA configs that run FCP extremely well. Maybe those are the 6GB+ configs, I don't know, but again I'll push back on claims that FCP runs horribly on NVIDIA across the board.

If all you're going to use your Mac for is to run FCP, then sure, sell your GTX 980 so you don't have to deal with the lack of boot screens and all that.

Or how about thinking and having context? I am talking about FCP here. So when I talk about OpenCL, I am referring to WHATEVER FCP uses. Not every and any use of OpenCL.

Again, explain how the 7950 beats my GTX 980 when A) my 980 has more memory and B) my 980 has more clock/memory speed? It just shows that the 980 is not very good at OpenGL as the 7950.

And again, you keep referencing the Titan X card. That should not be in the comparison. That has 4x the memory as the 7950. This is like benchmarking two i5 processors and someone just starts talking about the i7 benchmarks. The Titan X does not relate here. It is a different class of video card. Again, it BLOWS my 980 (and even the brand spanking new 1080 in some areas) out of the water across the board. Of course it will perform well, even if it has poor OpenCL because it is a crazy fast card!

 

[koyoot]

FCPX is all about Compute capabilities of your GPU, that OpenCL can expose for the application. Maybe this information can help you work out why AMD GPU are mostly better in any comptue rendering than Nvidia GPUs.

The trend which is seen in compute benchmarks is this: compute rendering - AMD. Image analysing - Nvidia.
FCPX is the first part.

 

[Ethosik]

FCPX is all about Compute capabilities of your GPU, that OpenCL can expose for the application. Maybe this information can help you work out why AMD GPU are mostly better in any comptue rendering than Nvidia GPUs.

The trend which is seen in compute benchmarks is this: compute rendering - AMD. Image analysing - Nvidia.
FCPX is the first part.

Yes, thank you. So just the fact that the Titan X is a crazy fast card conceals this. It is just a shame that going from the AMD 5870 to the GTX 980 is such a poor performance increase in FCP. Whereas going from the 5870 to the 7950 is a much bigger boost in performance.

Take a look at this: https://compubench.com/result.jsp?b...PU=true&pu-ACC=true&arch-x86=true&base=device

Video Composition. The GTX 980 and AMD 7900 series are pretty much identical. That give evidence that the 7950 is better than my 980. You can see Page 1 has more AMD than NVIDIA. I thought everybody knew that AMD was better at compute than NVIDIA? But NO, since the Titan X performs better.

It looks like the Fury X at least matches the Titan X.

 

[h9826790]

Granted, the 7950 is getting too old and still very expensive for the Mac edition, so I do not want to get that one .

If you can go for the 980, then there is no need to buy the 7950 MAC EDITION.

1) If you prefer to buy MVC 980, then you can buy MVC R9 280X, which works better in FCPX and cheaper.

2) If you don't need the boot screen. Any 7950 should work OOTB in OSX without boot screen (with any OSX that can run the 980)

3) If you want the boot screen, you can flash the card by yourself. I bought a new R9 280 (same GPU as the 7950), about a year ago, $150 for a new card, not that expensive IMO.

https://forums.macrumors.com/thread...tem-information.1808938/page-14#post-21607156

 

[SoyCapitanSoyCapitan]

Titan X is a crazy fast card, in Windows. On OS X not so much. You have to use this web driver based on an ancient 346 builds that still only has beta support for Maxwell architecture (there are people around here who desperately try to dodge this for financial reasons. Do not trust them.) So the Maxwell GPU in general use is not significantly better than a 680 on OSX.

And then the OpenCL performance of these cards is around 20% lower on OSX than Windows. Run Luxmark 2 and 3 to look at there difference.

But FCPX is well optimised for AMD and there's no Windows version so you don't have to deal with that.

 

[Ethosik]

Titan X is a crazy fast card, in Windows. On OS X not so much. You have to use this web driver based on an ancient 346 builds that still only has beta support for Maxwell architecture (there are people around here who desperately try to dodge this for financial reasons. Do not trust them.) So the Maxwell GPU in general use is not significantly better than a 680 on OSX.

And then the OpenCL performance of these cards is around 20% lower on OSX than Windows. Run Luxmark 2 and 3 to look at there difference.

But FCPX is well optimised for AMD and there's no Windows version so you don't have to deal with that.

Um what? His Titan X is getting WAY better results than my 680.

 

[SoyCapitanSoyCapitan]

Um what? His Titan X is getting WAY better results than my 680.

Read the post again. Read it properly.

 

[i make movies]

You have to use this web driver based on an ancient 346 builds that still only has beta support for Maxwell architecture

The beta nvidia web drivers for the (also beta) macSierra are based on the 367 builds. Do you think there will be significant performance gains once both are finalized?

 

[SoyCapitanSoyCapitan]

The beta nvidia web drivers for the (also beta) macSierra are based on the 367 builds. Do you think there will be significant performance gains once both are finalized?

One can only hope at this point.

 

[Ethosik]

Read the post again. Read it properly.

Um, I did? You said:

Titan X is a crazy fast card, in Windows. On OS X not so much

Which is false. Might be less than windows, but it is still crazy fast even in OS X.

You also said:

So the Maxwell GPU in general use is not significantly better than a 680 on OSX.

But that is false. The Titan X is better than my 980, and it just smokes my 680 even in OS X.

 

[SoyCapitanSoyCapitan]

Um, I did? You said:

Titan X is a crazy fast card, in Windows. On OS X not so much

Which is false.

So the Maxwell GPU in general use is not significantly better than a 680 on OSX.

But that is false.

We have real world and synthetic benchmarks from Barefeats putting the the top Maxwell cards in the same ballpark as the 680 in OpenGL. Only OSX of course.

If your app really does require more VRAM then there will be an improvement over an older card, but still nothing compared to Windows.

We have genuine benchmarks that OpenCL performance is lower on OSX than it is in Windows.

Your method of just saying "things are crazy fast" is cool, very subjective.

Some of us prefer objectivity, hard numbers, facts, things that don't require delusion or belief.

 

[AidenShaw]

Some of us prefer objectivity, hard numbers, facts, things that don't require delusion or belief.

Is that you, Hillary?

 

[Ethosik]

We have real world and synthetic benchmarks from Barefeats putting the the top Maxwell cards in the same ballpark as the 680 in OpenGL. Only OSX of course.

If your app really does require more VRAM then there will be an improvement over an older card, but still nothing compared to Windows.

We have genuine benchmarks that OpenCL performance is lower on OSX than it is in Windows.

Your method of just saying "things are crazy fast" is cool, very subjective.

Some of us prefer objectivity, hard numbers, facts, things that don't require delusion or belief.

So the Titan X is not fast? How is it that Asgorath gets better performance out of his Titan X than my 980 in OS X with FCP? How do you explain that? Shouldn't it perform the same since it is the same speed as the 680?

What about these benchmarks? https://compubench.com/result.jsp?benchmark=compu15d

I can clearly see a difference in OS X upgrading from my 680 to my 980. FCP is not it though. Therefore, the 980 does perform better so the Titan X would perform better than my 980.

Did I say OS X performance is better than Windows? No. Your statements are also COMPLETELY false.

"Titan X is a crazy fast card, in Windows. On OS X not so much"

Um, does OS X cause the CUDA cores to be cut by half? Does it force the card to operate at half the memory bandwidth? Does OS X only allow the Titan X to use only 2-4 GB of VRAM?

"So the Maxwell GPU in general use is not significantly better than a 680 on OSX."

Also not true AT ALL. Does it prevent me from using more monitors? Will it prevent me from using higher resolution? What about support for 144 hz? It will not have ANY affect on the very low frames with OS animations?

But what you are saying about it not being good with OpenCL, isn't that what I have been saying all along? Why it is such a low performance boost going from my AMD 5870 to the GTX 980?

Oh and finally, check this:

http://barefeats.com/gtx980ti.html

Look at the FCPX export graph. Really? 33.56 seconds is the same as 84.63 now? 33.56 = 84.63? It looks like the site you were referencing shows that FCPX has better performance between even the 980 and the 680 (I don't see it on my system, but here are the 'facts').

Oh and the 980Ti gets 1186 in the OCEAN WAVE OpenCL test while the 680 gets 519.

The Titan X would certainly be at least matching the 980 results. So please, tell me how they ALL perform the same as the 680?

How are these numbers in the same ballpark?

 

[SoyCapitanSoyCapitan]

Lol Asgorath, that is someone one I do not read at all. Someone I have on ignore for giving irresponsible advice to professional users to buy overpriced cards running on a dud driver that screws up pro apps.

Why are you posting Compubench? Why are you posting FCP benchmark to disapprove what I said? You don't know how to interpret my post and then you post things out of context.

So I'll waste some of my time with you even though it's like reasoning with a scientologist.

I said the best Maxwell cards are not significantly better than a 680 in OSX in general use and OpenGL.

Look at the OpenGL results (Tombraider and Dirt 3)

http://barefeats.com/gtx980ti.html

Pretty bad, right? On Windows that 980ti would be twice as fast (or more) as the 680. On a Mac the performance is a dud because of that ancient 346 driver. How long has the driver been stuck there? Hmmm. And the quality of that driver is getting worse.

Then we move on to compute. I did not say the best Maxwell cards are performing like the 680 at compute and OpenCL in OSX. If you think that I said this it's because you're not smart and you did not read my instruction to you to read my post carefully.

Certainly the Maxwell cards perform better there, they have more compute and CUDA cores for a start. However, under Windows they are yet again significantly faster.

If a Windows user sees you saying the Titan X is performing 'crazy fast' on OSX with the results that you are showing, you will be laughed at. Go on Reddit, Tom's Hardware or EVGA forums and show them your numbers. You will be the king of comedy to them, especially when they find out how much you paid for that card running on a dud driver.

'I spent all this time and money on this crap. There's no way I'm accepting that I was wrong. ' - Brainwashing 101

Now I hope not to read your knee jerk reply. Try to be humble and learn something for once. Don't be a Melania, you have to think for yourself.

 

[h9826790]

If a 980 can finish the test in about 27s in a cMP. Then a TitanX in the following Hackintosh can finish in about 15s is very reasonable. Yes, AFAIK, that 14s test was done on a Hackintosh, not a Mac.

"Core i7-4790K @ 4GHz
16GB 1666MHz RAM
GeForce TITAN X
10.11.4 with 346.03.06f01
Samsung 830 SSD

BruceX 5K finished in 14.62 seconds."

However, I don't think other card's are that VRAM limiting. Because my dual 7950 only has 3GB effective VRAM, but I can finish the test in also ~15s.

 

[Ethosik]

Lol Asgorath, that is someone one I do not read at all. Someone I have on ignore for giving irresponsible advice to professional users to buy overpriced cards running on a dud driver that screws up pro apps.

Why are you posting Compubench? Why are you posting FCP benchmark to disapprove what I said? You don't know how to interpret my post and then you post things out of context.

So I'll waste some of my time with you even though it's like reasoning with a scientologist.

I said the best Maxwell cards are not significantly better than a 680 in OSX in general use and OpenGL.

Look at the OpenGL results (Tombraider and Dirt 3)

http://barefeats.com/gtx980ti.html

Pretty bad, right? On Windows that 980ti would be twice as fast (or more) as the 680. On a Mac the performance is a dud because of that ancient 346 driver. How long has the driver been stuck there? Hmmm. And the quality of that driver is getting worse.

Then we move on to compute. I did not say the best Maxwell cards are performing like the 680 at compute and OpenCL in OSX. If you think that I said this it's because you're not smart and you did not read my instruction to you to read my post carefully.

Certainly the Maxwell cards perform better there, they have more compute and CUDA cores for a start. However, under Windows they are yet again significantly faster.

If a Windows user sees you saying the Titan X is performing 'crazy fast' on OSX with the results that you are showing, you will be laughed at. Go on Reddit, Tom's Hardware or EVGA forums and show them your numbers. You will be the king of comedy to them, especially when they find out how much you paid for that card running on a dud driver.

'I spent all this time and money on this crap. There's no way I'm accepting that I was wrong. ' - Brainwashing 101

Now I hope not to read your knee jerk reply. Try to be humble and learn something for once. Don't be a Melania, you have to think for yourself.

What is wrong with you? Why am I mentioning FCP? Um, because that is the point of this topic? You can't just pick and choose the benchmarks and say X always performs better than Y. That same site you referenced shows FCPX MUCH better with the 980Ti than the 680. Again, why am I mentioning FCPX? Because THAT IS THE ENTIRE POINT OF THIS TOPIC! I did not bring up OpenGL!

How am I supposed to interpret your two quotes here?

"Titan X is a crazy fast card, in Windows. On OS X not so much"

You did NOT specify OpenGL in that statement.

"So the Maxwell GPU in general use is not significantly better than a 680 on OSX."

You did not specify OpenGL in that statement. General use is a broad statement. It is not just limited to OpenGL or one specific thing. That statement is false because there are numerous benefits between the 980 and the 680 in "general use"

Again, I did not bring up OpenGL, you did. This entire topic is about FCPX which is OpenCL. Why throw benchmarks about OpenGL and say "See"?

Sigh. We cannot compare the performance to Windows on this topic. FCPX is not available on Windows, so that comparison is pointless. I don't care if it runs better on Windows. I cannot run FCPX on Windows so it does not matter.

Of course cards are better on Windows, I am not disputing that fact. It is more related to Direct X than drivers, because this has been going on for a very long time. It is still bad that I must run WoW on medium to get a decent constant framerate where I can run it on Max on Windows.

If this topic was about Tomb Raider or Dirt performance in OS X, I would be agreeing with you 100%. Windows is ALWAYS better for gaming.

 

[koyoot]

Whiplash, what SoyCaptain said was: Titan X is fast under Windows, but is 20% slower under OSX compared to windows on the same machine.

Did you miss the part where I was asking for single-GPU results? If I had 2 TITAN Xs, I'm assuming I'd get better than 14 seconds.

I think you are putting too much faith in it. FCPX is not optimised for Nvidia hardware, it is optimised for OpenCL. That Dual D500 has 4.4 TFLOPs of compute power. Your Titan X has 6 TFLOPs, and still looses badly(14s vs 8.9s).

FCPX is all about compute performance exposed by OpenCL. There is no excuse that there was dual GPU setup and here was single GPU setup. Titan X is bottlenecked by CPUs, and software on OSX side.

 

[Ethosik]

Whiplash, what SoyCaptain said was: Titan X is fast under Windows, but is 20% slower under OSX compared to windows on the same machine.

I think you are putting too much faith in it. FCPX is not optimised for Nvidia hardware, it is optimised for OpenCL. That Dual D500 has 4.4 TFLOPs of compute power. Your Titan X has 6 TFLOPs, and still looses badly(14s vs 8.9s).

FCPX is all about compute performance exposed by OpenCL. There is no excuse that there was dual GPU setup and here was single GPU setup. Titan X is bottlenecked by CPUs, and software on OSX side.

Sigh, Windows should not even be in this discussion! How hard is that to understand? Can I get FCPX on Windows? That is the discussion here.

If the topic was Tomb Raider performance on a Titan X on Windows vs OS X, then all the windows comparisons would be relevant.

The fact that I cannot get FCPX on Windows makes all Windows discussion irrelevant here.

He also said it performs close to the same as the 680. But again, that is a generalization and even the site he mentions has a benchmark showing MORE than a 20% gain between the 680 and the 980 in FCPX.

"So the Maxwell GPU in general use is not significantly better than a 680 on OSX."

And again, isn't this what I have been saying? AMD cards are better at OpenCL that FPCX uses.

Only 2 out of the 5 benchmarks on the site SoyCaptain listed show his point. The other three shows a big boost between the 980 and the 680.

 

[Asgorath]

Not that SCSC will see this, but I've never ever recommended people buy a card, ever. I've simply objected to his blanket claims about Maxwell or NVIDIA in general.

I think you are putting too much faith in it. FCPX is not optimised for Nvidia hardware, it is optimised for OpenCL. That Dual D500 has 4.4 TFLOPs of compute power. Your Titan X has 6 TFLOPs, and still looses badly(14s vs 8.9s).

FCPX is all about compute performance exposed by OpenCL. There is no excuse that there was dual GPU setup and here was single GPU setup. Titan X is bottlenecked by CPUs, and software on OSX side.

BruceX is a test that measures how quickly you can transfer raw 5K video frames across the bus to the GPU, have the GPU render them, and then send the raw final frames back across the bus. Dispatching video frames to 2 GPUs is going to be significantly faster than dispatching them to 1 GPU, because the render time is nothing compared to the transfer time.

In my experience, Final Cut Pro is not a strenuous compute application, it's all about data transfer and management. If FCP did all their ProRes encode/decode on the GPU it'd be another story, but at least on the NVIDIA card's I've tested, it appears to all be done on the CPU.

Again, if you compare the TFLOPS of 2 GPUs and get results that don't match that comparison, it's a very strong indication that TFLOPS have nothing to do with the performance in the test and that the bottleneck is elsewhere.

 

[koyoot]

Have you tested it with AMD GPUs? What if what you are asking is done on the GPUs, particularly AMD GPUs, and is not on Nvidia?

 

[Asgorath]

Have you tested it with AMD GPUs? What if what you are asking is done on the GPUs, particularly AMD GPUs, and is not on Nvidia?

No, I don't own any modern AMD systems (I have the 5770 from my cMP around somewhere, but that's it). So sure, perhaps the video decode/encode is being done on the GPU on AMD, which makes it "optimized for AMD" as some have claimed. It would also explain why NVIDIA performs worse than expected in FCP.