Resolved Mojave Hybrid AMD/NVIDIA Dual GPU (Limited Success)-No Display via RX 580

[Reindeer_Games]

So I found something interesting-and fun, metal acceleration on both a GTX770 and RX580 and hardware acceleration as well as I know to test it in both Mojave 10.14.3 and Win10-just no display in Mojave via the RX580.

My NVIDIA GPU is a MVC flashed GTX 770 Founders Edition 2GB-and reports correctly as a 770 since I purchased in 2016 from eBay. Doubtful a MVC original-but the bios is locked reports as such; I got it from an east coast seller. Possible it was just re-sold to me though.

I've been told the Quadro 4000 hangs on boot if attempting to boot Mojave with a Polaris:

...if I select to boot into Mojave, it gets stuck at the progress bar, most the time around half full or 3/4 full. Mojave will boot even with just the Quadro 4000 in, but there are graphic glitches.

My AMD GPU is a Sapphire Pulse RX580 8GB standard dual bios OOTB, using the primary bios. I was very happy with this setup's overall perfomance in both macOS and Window but sorely missed Bootscreen and PhysX in Windows.

Use caution-you can still trigger shutdowns if only using the PCIe boosters (i.e., DO NOT attempt from the Aux Boosters ALONE as any rendering will likely trigger a shutdown). Have a plan to provide power before even considering dual GPU's of any configuration long-term.

After getting both a cards running self regulated with the GTX additional power via the Aux Boosters and a Powerlink, and the RX580 via a Pixlas mod. I had planned running the GTX 770 from the Pixlas mod, but found shutdowns were still triggered even under Bootcamp boot unless power limted in Afterburner in kernal mode. Returning the GTX 770 to the PCIe Aux Boosters and Powerlink, and routing direct power to the RX 580 from the PSU solder points fixed the need to limit the power and allows full rendering in both OS's. In my personal opinion-a single lane solution such the Quadro 4000 for Mac (Fermi but I am unsure of Metal capability) or flashed single lane Kepler would be far closer to the spec power envelope for a standard workstation application but this is what I have tested and worked in limited function.

The RX580 in slot 1 was my first attempt, which worked-but with the same limitations I have currently, but the RX580 would run hotter due to Chill engaging (but not getting hot enough to trigger the fans). The 580 in slot 1 idles at 42 deg C, but in slot 2 it will idle at 38 deg C (prob cause the GTX FE's fan is circulating right next to the RX580's heatsink). Also, the GTX required a stand-off when in slot 2 to prevent the RX's fan's from clipping the solder-point's on my GTX which does not have a backplate.

A secondary case and PSU could be warranted-but even without display in macOS, the RX580 is providing accelerated compute that is visible by both app and monitor which are the main reason most of us use them and is why I absolutely consider this a success.

In macOS:

10.13.6 (17G4015- Baseline):
1) Both cards functional including NVIDIA web-drivers engaged, CUDA, and compute via the GTX.
2) Display via both the RX 580 and GTX 770 using both Apple and NVIDIA 387.10.10.15.15.108 web-drivers; natively or using manufacturer's drivers when only one GPU is attempted to display at a time.
3) FCPX fully utilizes both cards for compute. RED 6K samples debayering adjustments behavior was acceptable to me-I don't deal with RED files regularly though.
4) Any attempt of dual GPU display output creates an immediate log-out of the OS (with either drivers engaged).

10.14.5:

5) GTX770 must be used as primary-DP is functional (but when swapped to RX580 is not).
6) CLing is reporting Metal acceleration in both GPU's
7) FCPX does not appear to use the RX580, but Compressor does as a booster GPU.
8) Heaven/Valley uses only the GTX770, but recognizes the RX580's presence (or I have just not been successful in engaging the RX580 yet).
9) I am hoping a second monitor may change #6/#7 (or future software updates?).
10) Attempts at dual GPU display output are recognized by the OS and are visible by the OS under the Display/Arrangement Tab, but the desktop rendered by the RX580 does not output.

In Win10:

10) Best to use the RX580 as primary-both cards are usable.
10) GTX770 functions perfectly as a dedicated PhysX/CUDA card when using RX580's DP.
12) Certain settings on each AMD/NVIDIA control panel are only accessible when a monitor is connected to each manufacturer's card, so PhysX must be set to the GTX in the NVIDIA Control Panel prior to swapping to RX580 for use in Windows. It only needs to be done once usually, unless you move the cards slot location.

13) Ensure your cards are properly powered by getting a second PSU or provide power in an adequate manor.

I have a mDP switch that I've re-purposed (previously was buggy at times when using w/ my MBP, so I don't recommend it by name brand), that's working well for now between GPU's. When setting up I was just swapping the DP cable by hand once booted to Windows between the GTX/RX where hot-swaps aren't a problem. They aren't a problem in macOS, but again-you just don't have a display via the RX580 in Mojave.

I am planning a Pixlas mod to eventually power the RX580, or secondary PSU for both (undecided)-but wanted to share this setups viability for those that already have those bases covered. I personally favor a secondary PSU for longer-term use, but this is certainly viable for my use in Testing.

 

[bsbeamer]

Is there a reason you want/need to run both GPUs simultaneously? Between the two cards, the RX580 8GB will have much better macOS performance in Mojave, assuming your GTX 770 FE 2GB is GK104 (basically same as GTX 680 Mac Edition 2GB).

You can try swapping PCIe slots to see if it impacts macOS priority. Used to act this way with mixed GPUs, unsure if it still does.

 

[Reindeer_Games]

Bootscreen access and the added benefit of CUDA and PhysX in Windows, basically. In windows the 770 doesn't even hit the 120w power limiter, so I'm thinking the boosters and my Powerlink should be fine for testing since it hardly ever hit peak in macOS. I would rather use the RX580 in macOS, but I haven't achieved display yet. But even with only using the 770 in macOS there are considerable perfomance pickups-just like if I had web-drivers engaged.; it's fast.

Thinking a secondary display should set it up perfectly, for my application-since it should engage both. I just may not have dual monitors in Mojave-which I'm ok with. Biggest drawback so far is I had to pull my USB-C card to gain full access for the second GPU. I'm exploring transplanting options-some of which we've discussed previously.

 

[bsbeamer]

There used to be an external PCIe expansion box with a PCIe switch (basically an A/B switch). Have not seen it in at least 10 years, but might be worth looking into for your specific use. Probably would also solve your external PSU dilemma rather easily.

If you have already tried swapping slots with GPU and still not getting display with RX580, the driver conflict with multiple GPUs from two manufacturers still exists in Mojave and has not been addressed. Guessing your GTX770 is flashed? Using a GTX680 EFI?

 

[Reindeer_Games]

Guessing your GTX770 is flashed? Using a GTX680 EFI?

Edited and moved additional data to #1 post.

Correct it is flashed. I'm using a mDP switch to swap GPU's routed to my Cinema Display which works well enough for now.

 

[Reindeer_Games]

There used to be an external PCIe expansion box with a PCIe switch (basically an A/B switch). Have not seen it in at least 10 years, but might be worth looking into for your specific use. Probably would also solve your external PSU dilemma rather easily.

Something like startergo listed in the PCIe Expansion thread and having a lane-splitting switch would be nice, but well above my limited budget for experimenting. My budget allows for more repurposing, such as a used micro-ATX case with an upgraded PSU and a couple high-end PCIe extenders. Additionally, this isn't a project I'm personally looking to scale-but others may find value in it. I can deal with a random $200 project, not $2K-LOL. But it will still take a few months to gather the parts unless I stumble across one already with at-least a semi-modular PSU being recycled.

I found a couple used TB3 eGPU enclosures for reasonable and was considering picking up as chassis to mod, so I'd have hardware to return back to eGPU enclosures in the future. Either way I choose, these GPU's seem to be working in-tandem, which I was pretty exited about.

I pretty much consider this resolved for my application (other than the transplanting and powering mods), but want to leave it open for those that have the same/similar GPU's and more available equipment if they'd like to attempt scaling.

 

[Reindeer_Games]

First two are previous shots, one in HS, and the other after upgrading to Mojave.

The second two are new shot's are in Mojave showing rendering on the RX580 while installed alongside the flashed GTX770-but it's only reporting as 1GB (likely a Novabench reporting error, macOS recognizes the full 8GB in System Information).

 

[h9826790]

View attachment 826691 View attachment 826692

New shot's showing rendering on the RX580 while installed alongside the flashed GTX770-but it's only reporting as 1GB (likely a Novabench reporting error, macOS recognizes the full 8GB in System Information).

Interesting, I didn't know NovaBench has 2017 version. But anyway, it seems the GPU test still very outdated and unable to measure anything.

I just download and run it with my 1080Ti. Same as yours, shows 1GB VRAM. And obviously the FPS is limited to the monitor refresh rate due to V-sync.

 

[Reindeer_Games]

Interesting, I didn't know NovaBench has 2017 version. But anyway, it seems the GPU test still very outdated and unable to measure anything.

I just download and run it with my 1080Ti. Same as yours, shows 1GB VRAM.

It's the current version to my knowledge, and I'm attempting to use any rendering programs to see which GPU macOS is choosing to use for rendering. Novabench happened to be the first one that natively and solely picked the RX580 in Mojave while only being displayed via the 770. I didn't care what it scored, only that is it uses the RX580 to do it-and it achieved full speed alongside my flashed 770 in Mojave is a bonus.

The pics just show the 770 reporting at start-screen with the 580 being used in the top-left of the render-screen in one, while the other displays RX580 scores while still reporting GTX770 as the display GPU. My GTX would score similar in to my current score in High Sierra and prior, but has Never scored above 30 FPS in Novabench in Mojave, until now. My historical scores support the drop in performance as a result in upgrading to Mojave, I am arguing that this is just as good as web-drivers for me if Apple gets it fully implemented across programs, IF we can only have display in macOS. Adding even more value of having the RX580 in Windows 10 already, which is improved from what I've seen.

Thanks-then I'll presume the 1GB is a Novabench limitation (if I recall my 770 was the same before dropping in the RX580).

I'm basically attempting a workaround because I wish I could have a 1070 or better-but Mojave has it's limitations I have to work within for now. And spec-wise these cards are close to perfect to pair considering I prefer 60 Hz refresh. As a GTX user I'm sure you understand that PhysX is a cherished asset of any program that utilizes it; the same goes for CUDA users. The only feature I am seeing not "working" is display output via the RX580 in macOS, and I believe that it possible for that to change if someone were willing to make it happen.



Side note, before all the CPU/RAM/OpenCL tests quit working in Novabench in macOS (for me) this machine scored 2220-2430's overall, mostly contributed from the CPU's-but it's got it's perks still. Above are shots reporting success with freely available drivers to anyone with the 770 rendering in Windows-again I don't care about score, I'm amazed it's all working together natively for the most part. Both render approximately the same frame rate in Windows Novabench of ~65-69 FPS on GPU test but the GTX wins marginally.

I am also power limiting these cards currently-so I am expecting some improvement still once I get them powered properly. I'm very interested if a GT120 would work in Mojave the same way since it computes better and peak at only 50w, although I have logged use of up to 25% GPU and 1GB VRAM when utilizing the as a PhysX card in Windows peaking at around 48% power usage. That's around 110w, and the GTX's draw 2:1 so ~73.6w on its booster; the RX peaked at the 70% limiter which is ~129.5w and they draw theoretically around 3:1 ratio which would be ~97.125w (well above spec and in danger zone).

The GTX has been pulled for daily use and both booster returned to the RX580 until I receive additional cables to finish routing my Pixlas mod preventing power-booster damage my backplate. So again, have your bases covered if attempting further testing. I'm pretty sure a Pulse 56 and Kepler GTX combo would get the performance I truly want though all while keeping all my favorite features in Bootcamp.

 

[Reindeer_Games]

I found a few new articles on Apples Developer site explaining assigning rendering engines. A bit above my head but I'm looking into it more:

https://developer.apple.com/documen...selection_and_fallback_for_graphics_rendering

 

[Reindeer_Games]

I had a little free time today and a salvaged PC NVIDIA card came into my possession- it was a 9400 GT and after dropping it in place of the flashed GTX 770 and booting into Mojave:



Display was only via the AMD but it leaves one question to be asked-why only Apple/EFI flashed chipset's trigger the AMD's non-display when all setups work fluidly in High Sierra (including the 9400 GT!)? I understand why the older chipset isn't recognized in Mojave, but this is odd-by any angle of attack.

 

[bsbeamer]

In HS did you have NVIDIA Web Drivers installed, or was this clean OS install?

May want to do a comparison between the hardware extensions in Mojave and HS (clean install) to see what was eliminated and if any NVIDIA products are/were listed in HS that are not in Mojave. If you did have Web Drivers installed, just eliminate any of the ones with Web at the end.

I’ve tried posting selected drivers to Mojave and re-writing for macOS default driver version number without success. Also tried re-enabling the Volta driver from the pulled release with the new driver/replacement set without success. There is something in the NVIDIA “master” files that allows/rejects. Would need to spend massive amounts of time digging into and just not worth it. The hackintosh install route would actually be much easier for this specifically.

FWIW, cards with 9400M have basically been blacklisted from Mojave. Unsure if they were also blocked from HS, but I’m sure code was cleaned up during the Metal requirement.

 

[Reindeer_Games]

Clean install with restore from Time Machine during initial setup for my recent HS build I threw together; and 10.13.6 (17G4015) works beautifully with or without 387.10.10.15.15.108 web-drivers engaged; with both the 9400 GT or GTX 770 individually, simultaneously, and alongside the RX 580.

Windows 10 had driver conflict with all three GPU's in use, no issues with either the 9400 GT or GTX 770 (not my end-goal but tested for due diligence in exploring single lane GPU architectures).

In regard to the 9400 GT (Tesla), I hadn't found any kext to suggest Tesla architecture would be supported in Mojave (but exploring use) so that I might possiblely use a Apple GT 120 (Tesla)/Quadro 4000 for Mac (Fermi) to run my secondaries I recently picked up while dedicating the RX 580 to the Cinema Display. I understand neither (Tesla) are Metal approved, but both 9400 GT and GTX 770 display Metal quite well that was rendered on the RX 580 in High Sierra.

Truth be told, if I had an unlimited budget-I would find a Titan V/GV100 and call it good with the HS/Volta drivers but that is not likely to happen in the next 5 years-LOL.

EDIT: corrected macOS build number.

Additional note: I'm only using the 9400 GT in the interim to run 1 secondary while additional cables are in transit, only noticeable difference in HS/Win10 is no PhysX in Windows (9400 GT has never been PhysX capable). Added pic of all third party kext in HS:



Works like a charm. I need a PC Kepler to test Mojave with to see if it computes but does not display, since drivers are obviously present.

 

[Reindeer_Games]

According to this thread this update gives a GOP boot screen to:
NVIDIA TITAN Series:
TITAN X (Maxwell), TITAN X (Pascal), TITAN XP

GeForce 10 Series:
GeForce GT 1030, GeForce GTX 1050, GTX 1050Ti, GTX 1060, GTX 1070, GTX 1070Ti, GTX 1080, GTX 1080Ti

GeForce 900 Series:
GeForce GTX 950, GTX 950Ti, GTX 960, GTX 970, GTX 980, GTX 980Ti

GeForce 700 Series:
GeForce GTX 745, GTX 750, GTX 750Ti

Quadro Series:
Quadro Maxwell and Pascal products may be impacted. For support and additional details, contact OEM/Channel partner. If further assistance is required, visit //support.nvidia.com

I wonder will this update give a boot screen at least with a boot loader like refind?

Oh this could be good and would open up options for a Windows compute card.

 

[Reindeer_Games]

Graphics card compatibility for Final Cut Pro X, Motion 5, and Compressor 4
Learn which built-in and aftermarket graphics cards work with Final Cut Pro X, Motion 5, and Compressor 4.



Graphics card requirements
Final Cut Pro X, Motion 5, and Compressor 4 require an OpenCL-capable graphics card or Intel HD Graphics 3000 or later and 256MB of VRAM. If you're editing 4K video, using 3D titles, or editing 360º video, 1GB of VRAM is recommended.


"Supported built-in graphics cards
The following graphics cards come with Mac computers and support OpenCL in macOS Sierra, macOS High Sierra, and macOS Mojave.

• GeForce 320M
• GeForce 9400M
• GeForce 8600M GT 128MB*
• GeForce 8600M GT 256MB
• GeForce 8600M GT 512MB
• GeForce 8800 GS
• GeForce 8800 GT
• GeForce 9600M GT
• GeForce GT 120
• GeForce GT 130
• GeForce GT 330M
• GeForce GT 640M
• GeForce GT 650M
• GeForce GT 750M
• GeForce GT 755M
• GeForce GTX 660M
• GeForce GTX 675M
• GeForce GTX 680MX
• GeForce GTX 775M
• GeForce GTX 780M
• Quadro FX 5600"

https://support.apple.com/en-us/HT202239#nvidia



I'm waiting for this list to be finished (and maybe added to...) since it was "Published Date: November 15, 2018".

Or maybe they should at least word it more accurately-but I'd take completed +1/2 any day of week.

 

[bsbeamer]

I'm waiting for this list to be finished (and maybe added to...) since it was "Published Date: November 15, 2018".

Unsure what benefit this provides, or what you are hoping to be compiled. There is no GPU newer than a 2010 released model on this list.

The aftermarket FCPX list MAY be more helpful. Several of these GPUs do not work with Mojave:

Supported aftermarket graphics cards
The following aftermarket graphics cards support OpenCL in OS X Yosemite:

• AMD Radeon HD 7950 Mac Edition
• NVIDIA GeForce GTX 285 Mac Edition
• NVIDIA GeForce GTX 680 Mac Edition
• NVIDIA Quadro 4000 for Mac
• NVIDIA Quadro FX 4800 for Mac
• NVIDIA Quadro K5000 for Mac

 

[Reindeer_Games]

I'm pointing out that they have a marketing/reality conflict that should be fixed, and I was really hoping the 10.14.4 release was going to add support for GPU's that Apple sold and continues to advertise on their own webpage work with FCPX in Mojave, when my Tesla chipset sure isn't recognized in 10.14 at all (due to no drivers being present and thats why it's odd Apple's webpage says they are supported).

Metal isn't what drives Final Cut Pro X-my post points out it's OpenCL.

 

[bsbeamer]

FCPX (still) has several problems. Believe Optical Flow is still the only Metal2 supported feature/plugin. Apple also has not really updated their support/release notes for FCPX or the system requirements (https://www.apple.com/final-cut-pro/specs/), however they clearly mention RX 580 there.

Personally find Adobe Premiere Pro or Blackmagic Davinci Resolve perform much better when it comes to taking advantage of GPU. Even Avid Media Composer seems to be better than FCPX at times.

 

[tsialex]

I'm pointing out that they have a marketing/reality conflict that should be fixed, and I was really hoping the 10.14.4 release was going to add support for GPU's that Apple sold and continues to advertise on their own webpage work with FCPX in Mojave, when my Tesla chipset sure isn't recognized in 10.14 at all (due to no drivers being present and thats why it's odd Apple's webpage says they are supported).

Metal isn't what drives Final Cut Pro X-my post points out it's OpenCL.

Even if FCP-X with 10.13.6 support lots of GPUs, the moment you upgrade to Mojave the baseline is the Mojave GPU supported list. To think that Apple will backtrack from this is just unreasonable.

 

[Reindeer_Games]

Even if FCP-X with 10.13.6 support lots of GPUs, the moment you upgrade to Mojave the baseline is the Mojave GPU supported list. To think that Apple will backtrack from this is just unreasonable.

I'm just quoting their website-I agree to an extent, especially about 10.13.6; it works flawless.

But that doesn't change the fact that Apple's most premier app's are OpenCL based, and I never said go backwards since these cards are all capable OpenCL compute booster cards and additional secondary outputs; especially the Tesla's.

FCPX (still) has several problems. Believe Optical Flow is still the only Metal2 supported feature/plugin. Apple also has not really updated their support/release notes for FCPX or the system requirements (https://www.apple.com/final-cut-pro/specs/), however they clearly mention RX 580 there.

Personally find Adobe Premiere Pro or Blackmagic Davinci Resolve perform much better when it comes to taking advantage of GPU. Even Avid Media Composer seems to be better than FCPX at times.

Actually optical flow has worked great for at least 3 years when I started using it. I have a few friends that prefer DaVinci or Phase One, but I don't swim that deep in those waters; I just help out were I can and they deal with medium formats. I had Adobe for a couple years but was more than I really needed.

Yes, the RX 580 is listed-does great, and with adjustments handles most things I need. Powerful single lane card's aren't cheap though and there are only so many lanes available in a machine-and compute cards don't need to be powerhouses to be beneficial.

 

[tsialex]

I'm just quoting their website-I agree to an extent, especially about 10.13.6; it works flawless.

But that doesn't change the fact that Apple's most premier app's are OpenCL based, and I never said go backwards since these cards are all capable OpenCL compute booster cards and additional secondary outputs; especially the Tesla's.

FCP-X support material reflects that the app run with macOS versions post Sierra. You are limited with the GPU support for each of these macOS releases - not the contrary like you are implying.

 

[Reindeer_Games]

I'm not the one that wrote the statement that says all those GPU chipsets are supported by FCPX in all those OS's-Apples Marketing Department was the author; I just copy and paste.

I am well aware of what they mean- but it makes it hard to trust when not even their marketing material is correctly reflecting the operational reality.

 

[tsialex]

I'm not the one that wrote the statement that says all those GPU chipsets are supported by FCPX in all those OS's-Apples Marketing Department was the author; I just copy and paste.

I am well aware of what they mean- but it makes it hard to trust when not even their marketing material is correctly reflecting the operational reality.

Why you are trying to make this statement valid if you already know that the Mojave GPU support list trumps the FCP support list?

Let me quote you exactly:

"I was really hoping the 10.14.4 release was going to add support for GPU's that Apple sold and continues to advertise on their own webpage work with FCPX in Mojave, when my Tesla chipset sure isn't recognized in 10.14 at all (due to no drivers being present and thats why it's odd Apple's webpage says they are supported)."

It's reasonable to think that Apple should edit the FCP support page to make this clearer, but it's unreasonable to think that Apple will change the Mojave GPU support list to make the FCP GPU support list valid for Mojave too.

 

[bsbeamer]

Apple does not regularly update their FCPX or any “pro” software marketing materials. I’d argue this is one of the several reasons why many “pros” no longer use this software. They are terrible with release notes across the board.

NAB 2019 is in a few weeks. If you’re looking for additional Metal2 implementations, wait until then. Version 10.5 or 11 should be in the works for the transition to macOS 10.15, which will likely abandon OpenCL if the Mojave depreciation was any indication. They are also abandoning all 32-bit codecs and all QT7 based workflows. FCPX will need to be (re)written to support that in some portion.

Mojave has been a rocky transition at best, further complicated by the continued MacPro5,1 support. Stick with what “works” or else stick with High Sierra or previous.

 

[Reindeer_Games]

and thats why it's odd Apple's webpage says they are supported)

Because according to Apple-those driver should be present, but are not; I am presuming that is why it does not display. A mistake on my part in an over-statement.

Or maybe they should at least word it more accurately-but I'd take completed +1/2 any day of week.

Very reasonable-and wishful thinking that someone might hear how desperate users are to keep bootscreen.