Sample set up for CUDA programming for machine learning and gaming on macOS using a NVIDIA eGPU. Includes references, tutorials and generalizations that will apply to most hardware.
Important notice: as of 2020, the last compatible versions are macOS High Sierra (10.13) and NVIDIA CUDA 10.2. For more information, see eGPU on macOS Mojave and up.
- Guide on running NVIDIA eGPUs (with CUDA) on macOS
The ports/cables/adapters listed are simply the ones from the reference rig.
You can probably mix and match hardware in n→∞ different ways using the very same (or similar) software steps.
Just make sure to have a reasonable idea about your likelihood of success before you go on a buying spree.
Tip: check eGPU.io's for other reference implementations, Linux, macOS, Windows (pure or bootcamp) all included. Wealthy source of information, very helpful folks.
Thunderbolt is a multi-purpose interface designed by Apple and Intel. It can transfer a data at high speed, thus is adequate for transferring data between your computer and (e)GPU.
Long story short about Thunderbolt versions:
- Thunderbolt 3 looks like a USB-C port and provides 40Gbit/s bandwidth and is what you have in the latest computers, as the MacBook Pro from 2016-2017—or the Dell XPS, Razer Blade Stealth, etc.
- Thunderbolt 2 looks like a Mini DisplayPort port and provides 20Gbit/s bandwidth and is what you have in the previous generation computers, as the MacBook Pro from 2015 and earlier.
- Thunderbolt 1 is very outdated (and slow) and not worth taking into consideration anymore for these purposes.
The sample setup depicted in this guide is not optimal, given its need of Thunderbolt 3 -> Thunderbolt 2 conversion. This conversion is necessary in this case since the laptop has a TB2 port and the eGPU enclosure has a TB3 port. For this reason you end up requiring additional cabling and possibly limiting throughput to TB2 bandwidth.
If you are curious about the performance drops for GPUs connected via Thunderbolt (versions) versus PCIexpress, see a sample comparison. As usual, your mileage may vary and take things with a grain of salt.
TL;DR on Thunderbolt:
- If you are not restricted to TB2 because of an older laptop, go for a TB3->TB3 solution.
- If you are restricted to TB2, you can choose to go TB2->TB2 (cheaper) or TB3->TB2 (upgradable in the long term).
- Apple notebook as Apple MacBook Pro 13 inch 2015 (2x Thunderbolt 2)
- eGPU as Akitio Node (1x Thunderbolt 3, 1x PCIe), other options can be checked at eGPU.io buyer's guide.
- NVIDIA Pascal GPUs as GTX 1080 Ti (1x Displayport, 1x PCIe + others), Maxwell work as well.
- External display as Dell UP3216Q (1x Displayport + others)
- Displayport cable (2x Displayport, comes with display)
- Thunderbolt 3 cable (2x Thunderbolt 3, comes with Node)
- Apple Thunderbolt 2 cable (2x Thunderbolt 2)
- Apple Thunderbolt 2<->3 adapter (1x Thunderbolt 2 + 1x Thunderbolt 3)
- macOS native installation:
- macOS SIP disabling
- NVIDIA Web drivers, just match your OS version with driver version.
- NVIDIA CUDA drivers
- cuDNN
- Please refer to the Step-by-step Tutorials.
- eGPU enabling & automation:
- If on macOS 13.1 (High Sierra or maybe later): NVIDIA eGPU v2
- If on macOS 12.6 (Sierra) or earlier: Automate eGPU script
- WINE: for running Windows applications and games on macOS (or Linux) with a thin translation layer.
There are many tools which make the process of using WINE more straightforward and scalable; from the simplest to the most DYI:
- PortingKit: provides an extensive game library of tested "recipes" that help you creating standalone app wrappers; based on Wineskin Winery.
- CrossOver: commercial product by the maintainer of the WINE project, provides "recipes" for many applications and games; more robust and complete than PortingKit.
- WineBottler: a free, smaller and less maintained version that follows a concept akin to Crossover.
- Wineskin Winery: allows you to create standalone app wrappers with a high degree of customization; prefer the first port
- In-depth DIY guide of WINE on macOS: if you prefer to do it your way or want to know things more in depth.
- Parallels Desktop: Virtualization (VM) option; less optimal for gaming than WINE.
These tutorials are meant to be very (one could say overly) descriptive. Nonetheless, for reasons unknown, you may eventually find some slight variations. I will do my best to keep this updated with the intricacies of the process as people report it.
Fair to assume at this point that macOS 10.14 and over will never be supported by NVIDIA CUDA, as NVIDIA and Apple got into a deadlock. CUDA 10.2 is the last to support macOS up to 10.13. Please refer to macOS 10.14 support issue.
Support has improved substantially in the last months since macOS 10.13.4, both because of Apple and the eGPU community.
Currently the best alternatives are:
Detailed step-by-step guides will be provided over time as I am able to test them.
- Install macOS High Sierra (works with v13.1–v13.3). Please note v13.4–v13.5+ support is still experimental.
- Check if macOS System Integrity Protection (SIP) is enabled and/or enable it:
- Boot the computer in recovery mode: press and hold
Command⌘ + Rwhen hearing the chime sound. - Open the
Terminalapplication from the top menu. - Type
csrutil statusand pressEnter↩︎to see if SIP is enabled. - If SIP is not enabled, type
csrutil enable, pressEnter↩︎and reboot. - Reboot.
- Boot the computer in recovery mode: press and hold
- Apply the High Sierra supplemental update (if necessary).
- Install the corresponding [NVIDIA Web Driver](NVIDIA Web drivers to your OS version (remember, SIP must be enabled here).
- Reboot and log in normally.
- Disable macOS System Integrity Protection (SIP):
- Boot the computer in recovery mode: press and hold
Command⌘ + Rwhen hearing the chime sound. - Open the
Terminalapplication from the top menu. - Type
csrutil statusand pressEnter↩︎to see if SIP is enabled. - If SIP is enabled, type
csrutil disable, pressEnter↩︎ - Reboot.
- Boot the computer in recovery mode: press and hold
- Download and install the NVIDIA eGPU v2 package (if for whatever reason v2 does not work for you even after these exact steps, try v1).
- Shut down your computer.
- Now boot without the eGPU connected, log in normally, and again shut down your computer.
- Connect your eGPU enclosure and boot your computer:
- In most cases it should work with the eGPU connected before turning the computer on.
- In case you see an indefinite black screen, only plug the eGPU after the chime sound, just when the Apple logo shows up.
- Voilà!