MacPro 5,1 HD7970 - Bad Rendering Performance in FCP X 10.4.4

[alchemistics]

Dear Fellows

I recently upgraded my MacPro 5,1 (2 x 2.93 GHz 6-Core, 64GB RAM, 512GB 970NVME) with a Radeon 7970 3GB VRAM.

The card performs very good, Geekbench Score is at 100'000. Ungine Heaven @ extreme gets me 827 Points.
So far so good. Around twice the performance from my MacBook Pro (15" 2017 560). The anomaly I'm experiencing in FCPX is render times.

BruceX takes >60 Seconds and in the 4K Projects I'm working on the export time (ProRes) takes longer then my MacBookPro. Even a 7950 should get around 30s. Clearly something is wrong here. The card gets utilized fully according to iStat Pro.

I flashed the card successfully myself using netkas method. But behaves exactly the same with the original bios (over the switch).
I'm working on a clean Mojave (10.14.2) Install with FCP (10.4.4) on a Samsung 970 (~1500mb Read & Write).

I used to have this card in a Hackintosh with similar single core (cpu) performance and was getting BruceX export Times @ around 20 Seconds.

It's not only about the benchmark score. While working in FCPX the card seems to be performing way under its capability.

Does anyone have an idea where to start debugging?
I'm happy for any advice.

 

[fhturner]

Quick question: do you have any other GPUs installed along w/ the 7970?

 

[alchemistics]

Quick question: do you have any other GPUs installed along w/ the 7970?

Thanks for your answer.

I see your point. No there is just this one GPU installed in PCI Slot 1. Slot 2 is empty. Slot 3 is a Delock M.2 Adapter. Slot 4 the Sonnet USB-C card.

To me it seems like a FCP issue cause other application do get the right acceleration.
But I don't see how this can happen..

 

[joema2]

I don't think the Radeon 7970 supports Metal. I think Mojave requires a Metal-capable GPU. This was discussed here and in other MacRumors threads: https://www.pro-tools-expert.com/pr...-for-cheese-grater-apple-mac-pro-51-computers

If you could go back to an earlier version of macOS, there might be a better chance of it working. When the card worked in your Hackintosh, what version of macOS and FCPX was that?

This more likely legacy Mac Pro issue not a video issue (you're just seeing it manifest in video). You'll probably get a more informed answer on the Mac Pro forum.

 

[alchemistics]

I don't think the Radeon 7970 supports Metal. I think Mojave requires a Metal-capable GPU. This was discussed here and in other MacRumors threads: https://www.pro-tools-expert.com/pr...-for-cheese-grater-apple-mac-pro-51-computers

If you could go back to an earlier version of macOS, there might be a better chance of it working. When the card worked in your Hackintosh, what version of macOS and FCPX was that?

This more likely legacy Mac Pro issue not a video issue (you're just seeing it manifest in video). You'll probably get a more informed answer on the Mac Pro forum.

The card is fully metal compatible. Other (OpenCl) Applications perform as expected. Therefore this seems to be a FCP X issue. Any thoughts on this?

 

[joema2]

The card is fully metal compatible. Other (OpenCl) Applications perform as expected. Therefore this seems to be a FCP X issue. Any thoughts on this?

Even though that GPU is not on Apple's compatibility list, I know some people have reflashed it and it seems to work. The question is how broadly is that tested and does it cause issues for certain apps.

If you hold down OPT while clicking on System Information, then look under Graphics/Displays, does it say Metal is supported on that GPU? If it says "supported", does it say what feature set, such as
macOS GPUFamily2 v1?

When the app queries macOS, it's supposed to return a specific Metal feature set capability, not just yes or no. Without this the app doesn't know what features to use or how it might behave. It's possible without this info some apps (which use certain Metal features) might behave OK and other apps (which use other Metal features) might not: https://developer.apple.com/documentation/metal/mtlfeatureset

Ideally if specific feature set data is not returned, the app should throw an error and refuse to run or operate in a degraded mode. You can see how complex the varied Metal feature sets are in this table: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf

You could try to run GFXBench Metal, and see how that behaves, but there's no guarantee it's testing the same Metal features that FCPX 10.4.4 is using. My results are attached for both my 10-core Vega64 iMac Pro and my 2017 iMac 27 with RP 580.

We know both FCPX and Mojave work well with a variety of AMD hardware. Formerly your HD7970 GPU also worked better with FCPX, presumably an earlier version and before Mojave?

Unfortunately there is no config option in FCPX 10.4.4 or Mojave which can be adjusted to use certain graphics APIs. You could try a different GPU which is on Apple's supported list and see how that works: https://support.apple.com/en-us/HT208898

 

[alchemistics]

Even though that GPU is not on Apple's compatibility list, I know some people have reflashed it and it seems to work. The question is how broadly is that tested and does it cause issues for certain apps.

If you hold down OPT while clicking on System Information, then look under Graphics/Displays, does it say Metal is supported on that GPU? If it says "supported", does it say what feature set, such as
macOS GPUFamily2 v1?

When the app queries macOS, it's supposed to return a specific Metal feature set capability, not just yes or no. Without this the app doesn't know what features to use or how it might behave. It's possible without this info some apps (which use certain Metal features) might behave OK and other apps (which use other Metal features) might not: https://developer.apple.com/documentation/metal/mtlfeatureset

Ideally if specific feature set data is not returned, the app should throw an error and refuse to run or operate in a degraded mode. You can see how complex the varied Metal feature sets are in this table: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf

You could try to run GFXBench Metal, and see how that behaves, but there's no guarantee it's testing the same Metal features that FCPX 10.4.4 is using. My results are attached for both my 10-core Vega64 iMac Pro and my 2017 iMac 27 with RP 580.

We know both FCPX and Mojave work well with a variety of AMD hardware. Formerly your HD7970 GPU also worked better with FCPX, presumably an earlier version and before Mojave?

Unfortunately there is no config option in FCPX 10.4.4 or Mojave which can be adjusted to use certain graphics APIs. You could try a different GPU which is on Apple's supported list and see how that works: https://support.apple.com/en-us/HT208898

Hi @joema2

First, I highly appreciate your detailed response and knowledge. Thanks for this.

You see the System-Profiler Screenshot attached. It just says "supported".

I couldn't finish the whole GFX Metal Bench yet, but here are the partial results. I do find the "Feature Set" value there with "MTLFeatureSet_OSX_GPUFamily1_v2".

To isolate the issue I think there is no way around installing HighSierra and check FCPX performance there. I'll do this and report back.

In any case, a Vega 56 or 64 are valid options. Is there any (Non-Reference cooler) model, that won't block another PCI Slot (I need all of them)? Any way to get bootscreen on them? I'm awaiting the delivery of a GC-Titan-Ridge card to drive an LG5K UltraFine over two mDP connections.

Additionally, does it make any sense to wait for a Radeon 7nm card to install in the 5,1 MacPro?

So far, many thanks.

 

[sevoneone]

Question: Have you tried reinstalling FCP since swapping the card? Or at the very least trash your preferences after making a backup?

 

[joema2]

..I couldn't finish the whole GFX Metal Bench yet, but here are the partial results. I do find the "Feature Set" value there with "MTLFeatureSet_OSX_GPUFamily1_v2".

Your numbers look very good. The Metal feature set looks complete. I would expect no performance problems. BruceX is almost a pure GPU test. I cannot explain why it's so slow on your machine. Here are my BruceX times (unrendered TL, ProRes 422 output, recreate lib each pass, restart FCPX each pass):

12-core D700 Mac Pro: 17.0 sec
2017 iMac 27, RP 580: 15.8 sec
2016 MBP i7, RP 460: 36.2 sec
2017 10-core Vega64 iMac Pro: 13.5 sec

...To isolate the issue I think there is no way around installing HighSierra and check FCPX performance there. I'll do this and report back.

Regrettably that may be necessary. It is probably the best troubleshooting step.

...In any case, a Vega 56 or 64 are valid options. Is there any (Non-Reference cooler) model, that won't block another PCI Slot (I need all of them)? Any way to get bootscreen on them? I'm awaiting the delivery of a GC-Titan-Ridge card to drive an LG5K UltraFine over two mDP connections.

Sorry, not familiar with fitment issues inside a slot box; I have an iMac Pro. Likewise not familiar with the boot screen issue.

For playback or encoding H264 material, any Xeon-based Mac Pro will be at a disadvantage over a top-spec iMac. Xeon doesn't have Quick Sync. It's not uncommon for a MacBook Pro to beat a 12-core nMP D700 on H264.

Late-generation AMD cards have hardware transcoding called UVD/VCE. However this is totally separate from the graphics engine and all those GPU tests don't touch it. The only Mac that uses AMD's UVD/VCE is the iMac Pro. It works fairly well but it's still not as fast as Quick Sync on some H264 codecs.

If you are using all ProRes the above is irrelevant. But if your timeline is H264 and you are exporting to ProRes, it must still read and decode the H264 unless you created optimized media.

But the BruceX issue implies it's not related to video encode/decode, it's some other problem.

...Additionally, does it make any sense to wait for a Radeon 7nm card to install in the 5,1 MacPro?...

If your main use is video editing with FCPX, I'd say no. The only exception would be if Apple started supporting AMD's UVD/VCE acceleration on Mac Pros. That is impossible to predict but I'd be skeptical.

 

[alchemistics]

Question: Have you tried reinstalling FCP since swapping the card? Or at the very least trash your preferences after making a backup?

Thanks for the advice. Sadly the reinstallation process doesn't fix the problem.

Your numbers look very good. The Metal feature set looks complete. I would expect no performance problems. BruceX is almost a pure GPU test. I cannot explain why it's so slow on your machine. Here are my BruceX times (unrendered TL, ProRes 422 output, recreate lib each pass, restart FCPX each pass):

12-core D700 Mac Pro: 17.0 sec
2017 iMac 27, RP 580: 15.8 sec
2016 MBP i7, RP 460: 36.2 sec
2017 10-core Vega64 iMac Pro: 13.5 sec



Regrettably that may be necessary. It is probably the best troubleshooting step.



Sorry, not familiar with fitment issues inside a slot box; I have an iMac Pro. Likewise not familiar with the boot screen issue.

For playback or encoding H264 material, any Xeon-based Mac Pro will be at a disadvantage over a top-spec iMac. Xeon doesn't have Quick Sync. It's not uncommon for a MacBook Pro to beat a 12-core nMP D700 on H264.

Late-generation AMD cards have hardware transcoding called UVD/VCE. However this is totally separate from the graphics engine and all those GPU tests don't touch it. The only Mac that uses AMD's UVD/VCE is the iMac Pro. It works fairly well but it's still not as fast as Quick Sync on some H264 codecs.

If you are using all ProRes the above is irrelevant. But if your timeline is H264 and you are exporting to ProRes, it must still read and decode the H264 unless you created optimized media.

But the BruceX issue implies it's not related to video encode/decode, it's some other problem.



If your main use is video editing with FCPX, I'd say no. The only exception would be if Apple started supporting AMD's UVD/VCE acceleration on Mac Pros. That is impossible to predict but I'd be skeptical.

Thanks for the numbers. I attached the finished GFX Metal Benchmarks, the performance looks fine..

Trobuleshooting

I wen't through the troubleshooting process and installed a fresh copy of HighSierra to a spare internal SSD.
Installed a fresh FCP X, did the BruceX test and guess what: The rendering time went down to 36 second.

This result is more close to what it should be I guess. Over the clean HS I installed Mojave and did the test again. With almost same numbers (37s)!

So I thought the issue could be the internal PCI NVME drive. Next I cloned the whole system to the same internal Sata SSD where I had the better results on and bootet from that, did the testing and got the same bad number ~60s as on the NVME drive.

So the issue must be related to my specific system environment.

Therefore I did a clean Mojave install on the NVME drive and migrated my backup onto it.
Results: ~37s YES =)! It seems this has resolved it.

I'm still not quite sure how users of a 7950 / 7970 do get export times in the ~25s range. And why I got similar 20ish results with the same card in a Hackintosh with an inferior cpu..

FCP X Workflow

My current timeline is around ~60Minutes with H264 4K content in it. LUTS, Denoise etc. are applied which my MacBook Pro (560) can't handle without massive background-rendering. Additionally the current library / project can't remember background-rendering files anyways but this is another FCP X issue..

Even though the MacBook Pro does better encode and decoding, the timeline-performance seems to be smoother on the MacPro. Maybe the QuickSync does apply more to export rather than timeline and / or the 12 cores do make up for it. 64GB RAM, a desktop-GPU and better thermals surely are a benefit.

I do export to ProRes and H264 but mostly care about editing-workflow performance. Where the MacPro seems to shine even with H264 footage.
For now, the machine is performing fine again. I do look into upgrading to a Vega56 or the new Radeon7 in the future.

Thanks for your detailed and very helpful answers!

 

[alchemistics]

I eventually upgraded to a VEGA56.

BruceX is ~19 seconds now.

Details in this post.

 

[bronto11]

I eventually upgraded to a VEGA56.

BruceX is ~19 seconds now.

Details in this post.

How did you achieved this? I'm having similar issues with FCPX. It's not utilizing Vega 64 (Pixlas mod).
BruceX export takes ~40sec (Prores422 on PCI NVME drive).

 

[alchemistics]

How did you achieved this? I'm having similar issues with FCPX. It's not utilizing Vega 64 (Pixlas mod).
BruceX export takes ~40sec (Prores422 on PCI NVME drive).

I do know that performance decreased on 10.4.5, furthermore you should post other benchmark results to exclude a specific fcpx issue for efficient troubleshooting ..