Now that an eGPU can accelerate a Mac’s internal display, the question that naturally arises is how significantly the performance of the eGPU would be affected. I wanted to not only provide insight on this question but also show the performance of a graphics card inside a Mac desktop, the Mac Pro tower. This is the last modern Mac with full-length, internal PCIe slots.
The chosen graphics card for this comparison is an EVGA GTX 980 Ti Hybrid. It makes for a very nice pairing with the AKiTiO Node for an eGPU setup. The GTX 980 Ti Hybrid mounts inside the Mac Pro tower well with the only challenge coming from routing the liquid cooling lines.
The Mac Pro tower I tested is a 2010 model 5,1 with two 4x and two 16x PCIe 2.0 slots. In my opinion this is still the best Mac desktop available even though it’s six years old. The modular design makes upgrading components a joy. I’ve previously upgraded one of these Mac Pro towers to have 2x hex-core Xeon CPUs, GTX 980 Ti GPU, 128GB RAM, 512GB PCIe flash storage, and 4x 8TB Helium hard drives; all were mounted internally.
On the eGPU side, I’m testing the GTX 980 Ti Hybrid + AKiTiO Node with a Late 2016 13″ MacBook Pro through Thunderbolt 3 connection. For internal display acceleration, I’m using a headless ghost HDMI adapter. External display eGPU tests are through an HP 22″ monitor. The Unigine benchmarks ran at 1920 x 1080 fullscreen, max settings.
GTX 980 Ti Hybrid | Max PCIe Speed | Unigine Valley | Unigine Heaven |
---|---|---|---|
Mac Pro Internal PCIe Slot | 32Gbps-PCIe | 2,026 (48.4 FPS) | 1,658 (65.8 FPS) |
MacBook Pro External Display | 32Gbps-TB3 | 2,389 (57.1 FPS) | 1,756 (69.7 FPS) |
MacBook Pro Internal Display | 32Gbps-TB3 | 1,814 (43.4 FPS) | 1,084 (43.0 FPS) |
I’m sure you’re asking why the numbers from the Mac Pro internal 16x PCIe 2.0 slot are lower than External Display eGPU. Well, Apple crippled non-Mac GPUs to only be able to run at 2.5 Gbps lane speed. This gimpy connection effectively makes the graphics card run at 16x PCIe 1.1, which is close to 4x PCIe 3.0.
The important takeaway is the performance loss when you use an eGPU without an external display. Using an eGPU setup with an external display in macOS yields at least 30% higher performance than feeding that signal back into the internal display of your Mac.
It’s a considerable hit on performance. We all expected it though. Whether this trade-off is worth it depends on your situation and how you plan on using an eGPU. If you have questions about your external graphics card build, visit our forums.
[Nando] Q: How much bandwidth overhead does Optimus/X-connect accelerated internal LCD mode add?
Nando4 2019 update >> internal LCD mode is device-to-host traffic (incoming). Gaming and apps strain the host-to-device (outgoing) traffic. So feasible table results below are OK. Ignore the other waffle comparing bandwidths below since that isn’t applicable here.
As discussed on the eGPU.io forum, subtracting this value from our total link bandwidth, with a reference of how much bandwidth is needed for acceptable performance will then tell us if our eGPU is truly a portable desktop-replacement solution. One that doesn’t require an external LCD.
Resolution | Internal LCD overhead* (Gbps) | Theoretic remaining bandwidth for eGPU (Gbps) |
|||
32Gbps-TB3 | 16Gbps-TB2 | 10Gbps-TB1 | 4Gbps-EC2/mPCIe2 | ||
3840×2160 – 4k/UHD | 14.9 | 17.1 | 1.1 | unfeasible | unfeasible |
2880×1800 – 15″ Retina | 9.3 | 22.7 | 6.7 | 0.7 | unfeasible |
2560×1600 – 13″ Retina | 7.4 | 24.6 | 8.6 | 2.6 | unfeasible |
1920×1080 – FHD | 3.7 | 28.3 | 12.3 | 6.3 | 0.3 |
1680×1050 – WSXGA+ | 3.2 | 28.8 | 12.8 | 6.8 | 0.8 |
1366×768 – WXGA | 1.9 | 30.1 | 14.1 | 8.1 | 2.1 |
Lime: ~ 71% desktop GPU performance (x4 1.1 8Gbps or more bandwidth)
From this table comparison we see:
- x4 3.0 32Gbps-TB3 NVidia/AMD accelerated internal LCD traffic traffic overhead still leaves us with at least our desired x4 2.0 16Gbps or more bandwidth for pure GPU traffic.
- 16Gbps-TB2 with an external LCD to maximise it’s narrower PCIe link also gives ~85% desktop GPU performance levels. Internal LCD mode up to FHD can still deliver more than 71% desktop GPU performance.
- Slower 10Gbps-TB1, 4Gbps-EC2, 4Gbps-mPCIe2 links will see compromised performance, particularly at higher resolutions. Those require an external LCD attached, application of Frame Limiting tweaks and disabling of eGPU audio devices to maximise their narrow bandwidth for GPU traffic.
Related Articles
Share this Post
You can add my results too with fake display: 15′ MacBook Pro (Retina, Mid 2012, Thunderbolt 1) / Windows 10 / EVGA GeForce GTX 1060 SC 6GB / AKitio Thunder 2 External Display – 1376 (54.60 FPS) – Internal Display – 759 (30.1 FPS) – Internal Display – 1680×1050 (my Gaming resolution) – 1001 (39.7 FPS) – Here are another results from my friend without fake display (the laptop is with Intel Iris 6100 only): 13′ MacBook Pro (Retina, Early 2015) / Windows 10 / EVGA GeForce GTX 1060 SC 6GB / AKitio Thunder 2 External Display – 1246 (49.4… Read more »
Even though your results are in Windows, we can see a similar pattern in performance loss. It’s at least 30% difference when comparing internal vs. external display for eGPU setup.
Yes its at least 30% performance drop when we use the Internal Retina Display, but its better to use it because you have very good looking setup, without monitors etc on your desk!
If everyone is interested – to play all new and old heavy games on Very High/Ultra settings with 30-40 FPS right now you need around 1000 score on Unigine Heaven Benchmark. To play with 60 FPS You need at least 1200-1300 score and more!
Can we get the corresponding benchmark scores on win 10 to elucidate the driver difference?
Hi THEITSAGE, Putting these together have been great. I have a 2016 15″ on the way, as well as a Node and a GTX Titan X. I’ve had one lingering question that I haven’t gotten an answer to anywhere that might be of use for the kind of content you like to throw on the front page if it can be figured out: Clamshell mode. Many people use their Macbooks in Bookend docks or hengedock vertical docks in clamshell mode with an external monitor. Is there a way to run the eGPU in clamshell mode? I understand that you’ll probably… Read more »
I will definitely try InsomniaX with the Late 2016 MacBook Pro in Clamshell mode and report back. In theory the eGPU should remains active and functional as long as the Mac is not asleep while connected to the enclosure.
I’m not sure how that would work with a docking station such as the vertical Hengedock.
Awesome I appreciate you giving it a shot! With the Hengedock once it releases in Spring, yeah, I’m assuming I’d have to get everything going, close the laptop then insert it into the hengedock to connect to other peripherals. I’ll probably have to use the ports on the other (right) side for the eGPU.
Have any of you played with Thunderbolt based video capture devices to loop the output of the eGPU back in via a second port? This way we can overcome the 30% loss by having the game render to HDMI rather than a fake display. See below:
https://www.blackmagicdesign.com/sa/products/ultrastudiothunderbolt
I do acknowledge that the price is rather prohibitive, but it does overcome some of the biggest issues with portable gaming.
Very cool idea! Like you mentioned, not sure the added cost would be worth it for gaming though.
Hi Enjoy,
Do you know if I need to buy headless ghost if I use Windows via bootcamp ?
Will the performance of internal monitor be better with headless ghost/fake display ?
@NaYoung – No, the result will be the same!
Hi,
My laptop is Macbook pro late 2013 intel core I7-4850HQ 16GB ram.
What will be the best graphic card to get with Thunderbolt 2 box – Bootcamp setup?
So…does this performance penalty for using the internal display affect older, low-end cards that use less significantly less bandwidth? Say, for example a GTX 650?
Yes, there is performance penalty whenever using the accelerated internal LCD mode since display traffic needs to be routed back across the eGPU link.
As Pascal-based GPUs are fairly new to the whole eGPU scene do you think we’ll get improved performance over the next several months?
It’s a gradual improvement with Nvidia web drivers. At the moment, it’s rather buggy and the performance is not even close to Windows drivers.
Didn’t fully understand the 30% measured loss compared to the chart which seems to suggest that we should be at 85% of full speed (ie. around 15% performance loss).
Where is that extra 15% loss happening? Or did I misunderstand the chart?
Also would the 30% performance loss from outputting to the internal LCD be reduced by running games at lower resolution?
How much do the amount of internal RAM effect the performance on a system with a eGPU ? Im getting a Mac mini and dont wanna put too many RAM in the machine when I plan to get a eGPU TB3.
“The important takeaway is the performance loss when you use an eGPU without an external display. Using an eGPU setup with an external display in macOS yields at least 30% higher performance than feeding that signal back into the internal display of your Mac.”
Is the statement above still valid when using 2019 iMac and macOS 10.15.2?
@nando4, I have a question about the bandwidth requirement table calculated by Kramer’s Bandwidth calculator. It seems that calculator sums up 3 channels(RGB?) with 10bit each and ignore the Alpha channel which needs another bits. But shouldn’t it be 4 channels(RGBA?) with 8bit each than we get the 32bit true color?? For example, 1366×[email protected] 1366x768x60x10x3=1888358400bit/s≈1.9Gbps in the table. While I calculate it as 1366x768x60x8x4=2014248960bit/s≈2.0Gbps. PS: I didn’t post in the original disscusion in order not to resurrect that old thread. Besides, I guess the data compression difference seems huge between optimus and MS hybrid according to P106-090(Nvidia’s mining card but can… Read more »