MBP16,1 with RX 5700 XT via eGPU (Reports, Issues & Benchmarks)

[bsbeamer]

Been running some tests with an MBP16,1 configured with 2.4GHz, 5500M 8GB, 64GB RAM, 1TB SSD. Currently on 10.15.2.

Have a Sonnet eGFX Breakaway Box 650 GPU-650WOC-TB3, which is the OC model that provides 100W power delivery (newest model). Been running a Sapphire Pulse RX 580 8GB in that eGPU and testing an RX 5700 XT (model # 11293-01-20G) and has three DP connections and 1 HDMI connection on the card.

When running METAL tests, I've noticed horrible performance with the RX 5700 XT in Catalina compared to an RX 580 in Mojave. It's been 15-20% worse benchmark scores and around 33% slower in actual METAL usage. Have attributed this to poor drivers for the RX 5700 XT.

Had an opportunity to test some additional configurations with the MBP16,1 and RX 5700 XT in eGPU today and noticed some observations that might be helpful or interesting to some. All GB5 scores noted below are the "best" of around a three run average. Will run more another time. Whenever displays are connected, they are NOT set to mirror. Each are their own independent display and set to default/best for display mode. No crazy calibration profiles being used either, believe all were standard profiles for these tests.

25813 = 5500M dGPU with eGPU connected but not utilized for test
51591 = RX 5700 XT driving monitor via DisplayPort (27" ACD) and HDMI (Samsung 43" 4K TV)
27226 = RX 5700 XT driving monitor via DisplayPort (27" ACD)
27572 = RX 5700 XT driving monitor via HDMI (Samsung 43" 4K TV)
32905 = RX 5700 XT with no displays connected (headless)

That is not a typo for the DP & HDMI report, it's 51591 with two very different monitors connected. Ran that exact configuration 10+ times in various intervals and ALL have tested above 50,000 in METAL scores with GB5.

I attempted to shift DP connection/port being used, but does not appear to make a distinguishable difference so far. Cannot further test those today.

I REALLY do not like GB5 as a benchmark for METAL performance, but this clearly illustrates SOMETHING is wrong with the driver for RX 5700 XT. There is absolutely no explanation for this kind of "boost" with multiple displays vs. headless or solo displays.

I'm likely going to pickup some headless DP adapters to run additional tests in the coming weeks. Quick search and I'm not finding 4K@30Hz or 4K@60Hz for headless DP adapters. Some HDMI versions, but would prefer not to mess with three DP>HDMI dongles>HDMI headless for testing purposes.

 

[bsbeamer]

Further testing with some headless adapters for Sapphire Pulse RX 5700 XT 8GB via eGPU in GB5 METAL:

33785 = Driving DP1, DP2, HDMI @ Default for Display for Headless Adapters
32862 = Driving DP2, HDMI @ Default for Display for Headless Adapters

26550 = Driving HDMI @ Scaled UHD/4K for Headless Adapter

50424 = Driving DP1, HDMI @ Scaled UHD/4K for Headless Adapters
52428 = Driving DP2, HDMI @ Scaled UHD/4K for Headless Adapters

56822 = Driving DP1, DP2, HDMI @ Scaled UHD/4K for Headless Adapters

I cannot get DP1, DP2, HDMI, and DP3 all connected simultaneously. It's resulting in an instantaneous system freeze that requires hard shutdown and restart. If restarting with all connected, system freezes again upon loading the OS. Even when setting DP1, DP2, and HDMI resolutions to 1080 or below and then connecting DP3 system crashes.

There's an obvious METAL score benefit with multiple displays in the "correct" configuration.

Very preliminary initial testing shows METAL performance benefit with DP1/DP2/HDMI @ Scaled UHD/4K as well, but further testing is needed to actually confirm this.

Will test DP3 individually with monitors and adapters at another point. Too many system freezes to mess with for right now.

Will also try to test DP1, DP2 in combo and isolation. Then DP1, DP2, and DP3 if I can get DP3 to work.

 

[bsbeamer]

Using DP3 (port number 3) in any capacity on RX 5700 XT is resulting in system freezes or automatic restarts. Will not be further testing anything with DP3.

DP3 is the last/bottom port when mounted in the eGFX 650, furthest away from the bracket portion of the "L" on the card, closest to the PCIe pins. Port order for reference is DP1, DP2, HDMI, DP3.

 

[bsbeamer]

Pretty exhaustive test reports below in numerous configurations. Geekbench Version 5.1.0 (503054) METAL compute benchmark scores. There is a very clear benchmark benefit to running multiple monitors with HDMI @ 4K/UHD in some capacity. Triple monitors with DP1, DP2, and HDMI @ 4K/UHD results in the best scores.

RX 5700 XT with DP1, DP2, HDMI @ Default (Headless Adapters)
43494
43066
45155

RX 5700 XT with DP1, DP2, HDMI @ Scaled 4K/UHD (Headless Adapters)
70361
70890
66197

RX 5700 XT with DP1, DP2 @ Scaled 4K/UHD, HDMI @ Default (Headless Adapters)
44714
43112
43769

RX 5700 XT with DP2, HDMI @ Default (Headless Adapters)
42768
43468
43402

RX 5700 XT with DP2, HDMI @ Scaled 4K/UHD (Headless Adapters)
64374
64911
64522

RX 5700 XT with DP1, HDMI @ Default (Headless Adapters)
44706
42807
42660

RX 5700 XT with DP1, HDMI @ Scaled 4K/UHD (Headless Adapters)
65105
65916
65871

RX 5700 XT with DP1, DP2 @ Default (Headless Adapters)
43976
42612
43287

RX 5700 XT with DP1, DP2 @ Scaled 4K/UHD (Headless Adapters)
44703
43057
43747

RX 5700 XT with DP1 @ Default (Headless Adapter)
43784
45470
43389

RX 5700 XT with DP1 @ Scaled 4K/UHD (Headless Adapter)
43020
44824
42705

RX 5700 XT with DP2 @ Default (Headless Adapter)
45200
46740
43527

RX 5700 XT with DP2 @ Scaled 4K/UHD (Headless Adapter)
42810
43605
43759

RX 5700 XT with HDMI @ Default (Headless Adapter)
43537
43822
43614

RX 5700 XT with HDMI @ Scaled 4K/UHD (Headless Adapter)
33548
35634
34081

5500M (tested with eGPU connected)
29418
28296
28486

5500M (tested with nothing connected)
29060
29751
28401

 

[bsbeamer]

FYI, the Sapphire Pulse RX 5700 XT 8GB is specifically listed as macOS 10.15.1+ compatible with Breakaway Box 650 (OC) model by Sonnet on their eGPU list.

So this DP3 port issue is not a known or reported issue. Sonnet's offices are closed until January for the holiday and will reach out to them in the new year.

https://www.sonnettech.com/support/downloads/manuals/Compatibility_Graphics_Cards.pdf

 

[bsbeamer]

Adding this info. Some were looking for ROM details and EFI info.

Sapphire Pulse RX 5700 XT in Sonnet eGPU 650 OC:
Radeon RX 5700 XT:
Chipset Model: Radeon RX 5700 XT
Type: External GPU
PCIe Lane Width: x16
VRAM (Total): 8 GB
Vendor: AMD (0x1002)
Device ID: 0x731f
Revision ID: 0x00c1
ROM Revision: 113-1E4112U-O45
Automatic Graphics Switching: Supported
gMux Version: 5.0.0
Metal: Supported, feature set macOS GPUFamily2 v1
GPU is Removable: Yes

Native 5500M 8GB:
AMD Radeon Pro 5500M:
Chipset Model: AMD Radeon Pro 5500M
Type: GPU
Bus: PCIe
PCIe Lane Width: x16
VRAM (Total): 8 GB
Vendor: AMD (0x1002)
Device ID: 0x7340
Revision ID: 0x0040
ROM Revision: 113-D3220E-190
VBIOS Version: 113-D32206U1-019
Option ROM Version: 113-D32206U1-019
EFI Driver Version: 01.01.190
Automatic Graphics Switching: Supported
gMux Version: 5.0.0
Metal: Supported, feature set macOS GPUFamily2 v1

 

[jinnyman]

Pretty interesting. Wonderful job investigating unsettling problem!
I have no idea how connecting more monitor gives boost in performance. In theory, the bench should show the same level of performance regardless of how many monitors are connected? Well, to be exact, we should have slight (probably insignificant enough) performance hit due to increased number of pixels a graphic needs to drive.

 

[bsbeamer]

Sonnet support is stating this DP3 problem is an issue with "this specific card" and they have 5700 XT's working completely fine. They will not disclose those models or part numbers or manufacturers of the cards they have personally tested and are personally using. They're stating the issue is (obviously) not with the eGFX 650 OC so their support cannot do anything.

I do agree (mostly), however my gripe is this specific GPU is listed in their published compatibility list under part number 11293-01-20G:

https://www.sonnettech.com/support/downloads/manuals/Compatibility_Graphics_Cards.pdf

The M3 footnote just states 10.15.1+ is required:

Kudos to Sonnet support for responding (more than some companies even do), but it's a bit disingenuous to make that claim then basically pass the buck to Sapphire...

Sapphire is saying it could be a hardware problem and to RMA the card since it's within extended holiday return windows.

Dealing with the vendor who sold the GPU currently, but I'm likely going to be slapped with a restocking fee since this DP3 problem cannot be easily identified on an RMA bench. They'll plug the GPU in, see one of the outputs works fine and then will not issue the full refund. Been there, done that with computer hardware in the past. Weighing if it's even really worth pursuing further, or just wait until 10.15.3 to see if it is magically fixed.
[automerge]1578080385[/automerge]

Pretty interesting. Wonderful job investigating unsettling problem!
I have no idea how connecting more monitor gives boost in performance. In theory, the bench should show the same level of performance regardless of how many monitors are connected? Well, to be exact, we should have slight (probably insignificant enough) performance hit due to increased number of pixels a graphic needs to drive.

Several things are contributing here.

First, the macOS Catalina 10.15.2 drivers for RX 5700 XT are just not that good. Apple will need to at least address performance issues with W5700X and those improvements should trickle down. Hopefully we'll see 10.15.3 released in a few weeks, likely/hopefully around the same time W5700X orders start for MP7,1.

Second, Geekbench is just not a good GPU benchmark or a "great" benchmark in many other ways. It's inconsistent on identical machines and results are not repeatable even when running in intervals. Taking 5-10 results in intervals and averaging is really the only way to get something close to a reliable reference number. Simply too much work for the amount of testing I was doing here.

I'm speculating there is some "bonus" built into the scores for resolution(s) available. Can't figure out another way to explain the major differences.

In theory, using an eGPU to not drive any displays and JUST as a hardware acceleration device should yield the strongest results when it comes to METAL or any compute test. It's taking all the burden off the GPU and letting it JUST process compute. That has been far from reality in a number of situations on macOS going back to the XSERVE days.

 

[bsbeamer]

Received a message asking for GB5.1 OpenCL scores. These are with DP2 & HDMI connected to RX 5700 XT via Sonnet eGPU 650 OC (driving two monitors) and 5500M driving MBP16,1's display.

5700XT 8GB Metal: 65753
5700XT 8GB Open CL: 58646

5500M 8GB Metal: 29873
5500M 8GB OpenCL: 28437

 

[bsbeamer]

The Sapphire Pulse RX 5700 XT 8GB has a dual BIOS, with one of the modes being "silent" and supposedly it's totally based on the original stock RX 5700 XT reference BIOS without any modifications.

The GPU ships in standard mode, switch to the right when reading the words BIOS SWITCH upright. It can easily be switched with a small flathead screwdriver if you cannot easily reach. Honestly, neither BIOS is properly labeled in an easy to identify manner. The switch is on the power side of the card itself, towards the DP/HDMI ports, so should be accessible without physically removing the card for nearly all installations.

Silent mode makes very little audible sound difference in all METAL testing in macOS 10.15.2 and does NOTHING to address the DP3 issues mentioned above. For performance, there is around a 7-10% drop in almost everything vs. standard mode.

Cannot see this being worth it for many people looking at an eGPU unless it's literally to drive additional displays, but at that point this is probably not the right GPU for you anyway.

 

[bsbeamer]

More METAL and driver issues being observed with both 5500M and 5700 XT:

AMD Radeon RX 5700

Just a head's up for anyone considering RX 5700 or RX 5700 XT - there are pretty significant issues with the METAL driver for 5XXX series in Catalina 10.15.2 when used with Adobe software. It's particularly noticeable with the Mercury GPU Acceleration Engine in AE CC 2020 on many effects. It's...

forums.macrumors.com

 

[bsbeamer]

This will be the last report with 10.15.2 and will update MBP16,1 to 10.15.3 in the coming days.

Had a chance to run an identical Adobe Premiere Pro CC 2020 project on two machines for an overnight render. Project was fully cloned to an external SSD and run on both MBP16,1 (Catalina 10.15.2 with RX 5700 XT in eGPU) and MP5,1 (Mojave 10.14.6 with RX580). The same exact very intensive timeline render with multiple streams and effects present.

MBP16,1 = 12:59:24
MP5,1 = 13:53:06

MBP16,1 was nearly an hour faster
Timeline was 115035 frames (about 1 hour, 4 minutes)

MP5,1 is dual X5690 (2 x 3.46GHz) with 128GB RAM and 1TB NVMe boot
MBP16,1 is 2.4/64GB/1TB/5500M 8GB

 

[bsbeamer]

GB Version 5.1.0 (503054) METAL scores with 5500M 8GB without eGPU (native, nothing connected) with macOS 10.15.3 (19D76):

27199
27715
27040

Will attempt to test a few various configs with eGPU RX 5700 XT 8GB next week.

 

[bsbeamer]

METAL scores in GB Version 5.1.0 (503054)
MBP16,1 with RX 5700 XT in Sonnet eGFX 650 OC eGPU
macOS 10.15.3 (19D76)

DP1, DP2, HDMI:
65812
68384
72509
69332
69542

HDMI only:
35805
35633
35459

Exact same issues with DP3 as previous. System locked up almost immediately when trying. Not testing this anymore.

In all testing so far with 10.15.3, running into the same METAL issues within Adobe CC 2020 video products with SOME effects and plugins. Doing some further testing with those, but nothing indicating the issues have been addressed at all.

Does not appear anything with AMD 5XXX drivers has been addressed with this release.

Can now 100% confirm METAL issues are still present with 10.15.3 19D76 with AMD 5XXX in Adobe After Effects CC 2020 (17.0.2 Build 26) on some effects with both RX 5700 XT and 5500M.

Turning METAL off (Mercury Software Only) and the issues disappear immediately.

These issues are NOT present with RX580 in Mojave 10.14.6 18G3020 (latest update) in AE CC 2020 (17.0.2 Build 26) with METAL acceleration turned on (Mercury GPU Acceleration Metal).

 

[bsbeamer]

Had an opportunity to run through a day with the eGPU setup with only one display connected, BUT had two headless adapters set to mirror that same display.

Setup was DP1 with headless adapter, DP2 was driving display (27" ACD), HDMI with headless adapter. DP3 was not connected for the same reasons reported above.

All "displays" on eGPU were set to mirror to the best settings for the LED Cinema Display. Did take some tricking to get the HDMI to connect properly (reset resolution manually, then set to mirror, then change to optimize for LED Cinema Display and best for LED Cinema Display).

For whatever reason, this trick increases eGPU performance and seems to increase GPU acceleration with video work vs. standalone with single monitor (via DP1 or DP2). Working on trying to measure that more substantially vs. straight observation.

Some GB Version 5.1.0 (503054) scores for METAL:
66115
65621
64911

There are several threads on MBP and MacMini forums about internal ports not correctly working with eGPU setups and the mirror trick has been used as a workaround.



Personally had never set individual displays to mirror while others were not, but it is possible if you OPTION + CLICK/DRAG the display (so they stack) in the arrangement tab. You may need to OPTION + CLICK the resolution > scaled to bring up the hidden menu for additional resolution options. Straight clicking the checkbox for Mirror Displays makes all the displays the same.

The mirror display with headless adapter trick might be worth exploring for those with eGPU setups and only one or two external monitors, regardless of base machine.

 

[bsbeamer]

Was able to quickly test a Sapphire Pulse RX580 8GB with single monitor, with and without headless adapters.

Single monitor GB5 METAL scores:
38320
38570
37040

Single monitor with three headless adapters set to mirror GB5 METAL:
42727
42512
41739

Single monitor with two headless adapters set to mirror GB5 METAL:
41117
42238
42636

Worth noting - maxing out with 4 ports total was very slow performance on the monitor (even with all set to mirror). Just seemed slow to respond, almost like it was 1/2 rate. This was DP1, DP2, HDMI1, HDMI2. There's no way this would be usable in this configuration.

Dropping to 3 ports in use total and everything was normal. This was DP1, DP2, HDMI1.

The performance difference is not drastic at all with 3 headless vs. 2 headless, but there certainly is a "boost" with multiple monitors in some capacity.

Certainly would consider headless adapters if you're trying to increase GPU performance in any capacity.

 

[arturmees]

Hey. Got a razor core x with Sapphire Nitro + Radeon RX 5700 XT, 8GB, 15' MacBook Pro 2019...anybody had any issues with headless HDMI adapters only working once? I cannot get it to recognise them anymore..done the reset on SMC/NVRAM etc..cannot seem to find many answers to that from google? HDMI cable to TV 4k works ok, running a display over DP.. getting better GB5 results..but headless adapters are just toast..like got 3 of them, got two for eGPU and one for Mac mini..the moment I removed the HDMI headless adapter...it does not recognise it afterwards..cheers

 

[Dragos Teglas]

MacBook Pro (15-inch, 2019)
Processor: 2.3 GHz 8-Core Intel Core i9
Memory: 32 GB 2400 MHz DDR4

eGPU enclosure: Razer Core V2
GPU: AMD Radeon RX 5700 XT (Sapphire Built)

Geekbench 5 Results:
No Monitor in eGpu
Metal Score: 47 336

1x DP1 linked
Metal Score: 71 675

I'm assuming there is still some work to be done for improving these drivers. As we can see they have improved from December until now quite a bit.

 

[bsbeamer]

anybody had any issues with headless HDMI adapters only working once?

Some very cheap ones are finicky and can "burn" out when when receiving power. Usually buy the packs of three or more to avoid an issue and it's only few dollars more. Maybe try a different brand.

Do the adapters you've used in eGPU then work in MacMini?
Adapter that confirmed working in MacMini then work in eGPU?
Do your DP port(s) have any issue?

 

[bsbeamer]

Additional eGPU tests, in case they're useful or helpful for anyone. Needed to pickup an iMac for an indefinite work from home situation due to Coronavirus. Luckily got "on sale" and should be a stopgap desktop for at least a year or two.

Identical Sapphire Pulse RX 5700 XT 8GB in Sonnet eGFX 650 OC used on an iMac.

Machine specs:
iMac 27" 2019 Retina 5K
3.6 8-Core i9 with 128GB RAM and Vega48 dGPU
Catalina 10.15.3 19D76

GB5 METAL scores for RX 5700 XT with DP1 (headless), DP2 (headless), HDMI (TV/monitor):
52113
51597
50961
51247
51531

Regardless of what benchmarks say, this machine is straight up faster in many tasks vs. MBP16,1. The METAL scores absolutely do not do justice to just how much "better" the initial METAL tests are on this iMac vs. MBP16,1.

 

[ramonycajal]

Additional eGPU tests, in case they're useful or helpful for anyone.

Thanks for sharing all the benchmarks! I am new to eGPUs, and I'm looking to buy the RX 5700 XT/Razer Core X. I have a MBP16,1, and I would like to use the eGPU with the internal monitor in both macOS and bootcamp Win 10 (after hopefully overcoming all problems associated with Windows 10).

I’ve read through all your extensive benchmarks, but I can’t seem to find any test with no displays connected and headless adapters? (Sorry if I missed that)

Would adding headless adapters increase the eGPU performance, when using the internal display of the 16 inch MBP (in macOS and/or Win 10)? If so, how many/which headless adapters would you ideally add? Thanks

 

[bsbeamer]

eGPU without any display or any adapter from #1
32905 = RX 5700 XT with no displays connected (headless)

I have not tested the Razer Core X.

You can see the results above. Headless adapters increase eGPU performance tremendously for apps that can utilize METAL via eGPU. I'd suggest 2 DisplayPort and 1 HDMI. Mess with settings. The adapters are $2-5 each and even less when bought in packs or sets. Buy enough to fill every port and do some tests. I'd get at least 4 DP and 2 HDMI in case one is dead on arrival.

Your app will determine performance gain. Not all apps can use multiple GPUs. The MBP16,1 & eGPU will be multiple GPU - it's a dGPU and eGPU.

 

[ramonycajal]

eGPU without any display or any adapter from #1
32905 = RX 5700 XT with no displays connected (headless)

Thanks for the quick reply and help! So just to make sure that I am not misunderstanding anything: the score above (32905) is with the RX 5700 and no adapters? And if I were to add a few adapters (e.g. 2 DisplayPort and 1 HDMI), then I should expect an ever higher score than this?

 

[bsbeamer]

32905 = RX 5700 XT with no displays connected (headless)

Same post #1:
51591 = RX 5700 XT driving monitor via DisplayPort (27" ACD) and HDMI (Samsung 43" 4K TV)

Post #4 this setup yielded the best testing result numbers:
RX 5700 XT with DP1, DP2, HDMI @ Scaled 4K/UHD (Headless Adapters)
70361
70890
66197

Will reiterate, I've run into a ton of issues getting DP3 (the last DisplayPort on the RX 5700 XT card) to work in any capacity. I would not suggest utilizing. Running the last DP on RX580 was extremely slow monitor visual responsiveness. Great "numbers" but not even useable in real world in any capacity.

My personal opinion is 4 monitors (true or virtual via headless) is too many to drive off a single TB3 via eGPU. That is a personal opinion. It is possible and people do so with cards other than RX 5700 XT, but the real world usability in all testing was not worth the tradeoffs for my needs.

 

[ramonycajal]

32905 = RX 5700 XT with no displays connected (headless)

Same post #1:
51591 = RX 5700 XT driving monitor via DisplayPort (27" ACD) and HDMI (Samsung 43" 4K TV)

Post #4 this setup yielded the best testing result numbers:
RX 5700 XT with DP1, DP2, HDMI @ Scaled 4K/UHD (Headless Adapters)
70361
70890
66197

Will reiterate, I've run into a ton of issues getting DP3 (the last DisplayPort on the RX 5700 XT card) to work in any capacity. I would not suggest utilizing. Running the last DP on RX580 was extremely slow monitor visual responsiveness. Great "numbers" but not even useable in real world in any capacity.

My personal opinion is 4 monitors (true or virtual via headless) is too many to drive off a single TB3 via eGPU. That is a personal opinion. It is possible and people do so with cards other than RX 5700 XT, but the real world usability in all testing was not worth the tradeoffs for my needs.

Thanks a lot for explaining this so thoroughly. I think I might have been interpreting your benchmarks incorrectly, but reading your comment just now I think I finally get it.

For some reason I assumed that all your benchmarks only referred to actual monitors (except "32905 = RX 5700 XT with no displays connected (headless)"), and that the benchmark scores using monitors aren't applicable to their headless adapter counterparts (if that makes any sense).

I will buy some DisplayPort and HDMI adapters and start with from there. Thanks again for all the advice!