Cheap and Painless eGPU Thrills on a 2013 MacBook Pro

I would have certainly preferred to do this, but it would have involved a $2000 upgrade first! (Well, barring the TB3-to-TB2 adaptor approach, but I've seen almost no builds doing this.) Also, from what I've seen, there's a lot of uncertainty regarding eGPU support with TB3; many people have reported problems on TechInferno etc., and even AKiTiO is unclear about Mac support. (If you check the specs for the Node, you'll see that Macs are listed as unsupported. Does this mean OSX-only or BootCamp as well? Who knows?) Might be worth waiting a year for things to shake out.

Also, I was optimizing in part for cheapness. Truly great gaming cards cost $300 and up, and $400 total was about as much as I was willing to pay for an upgrade.

TB3 enclosures are also more expensive (and the affordable ones aren't readily available yet), have minimal performance gains[1] and have limited support on MacOS.

1 - https://youtu.be/vvAB3U5umug?t=187

> This is interesting academically, but I wouldn't buy a hacked-together TB2 enclosure.

Huh? The article is about a setup built using an AKiTiO PCIe box. You even mentioned that manufacturer as an alternative. I guess it's not officially "for eGPU use", but that's just a question of power, and a small SMPS brick is fine for that.

> You can plug a TB3 device into a TB2 laptop with a simple adapter. Apple sells theirs for $29; Chinese versions will be cheaper.

Apple's adapter has turned out not to work on Linux. It might not work with BootCamp either. Seems to require a special driver:

https://bugzilla.kernel.org/show_bug.cgi?id=189731

I spent a while trying to convince myself I could get away with a Mac Pro (Trashcan) hooked up to an external GPU to do VR & Deep Learning using a TB2 enclosure. After fighting for hours (days?) with how the PCI lanes were allocated on Windows, I finally got output to an external display, but it still ran below VR spec according to the various benchmarks. If you wanted to switch back and forth between Windows and Mac, it always required a few power cycles because Windows was constantly thinking it was reconfigured.

I returned the whole thing and bought a beefy gaming PC, which now has an absurd amount of power in it (a 1080 + 2x970... I had forgotten how much fun it is to continually upgrade a machine). For power on the go, I have a Razer Blade with a 1060 in it. It's a shame, because I prefer to develop on a Mac, but I need modern hardware ¯\_(ツ)_/¯.

> You can plug a TB3 device into a TB2 laptop with a simple adapter.

I don't know of any such adapter. Do they exist? The Apple one allows you to use TB2 devices with your new TB3 laptop, not the other way.

I don't understand, this is $300 for a box with a port adapter and a $35 power supply inside of it?

There is no need for 400 watts of psu for new generation nvidia cards.

Also, all the features you mentioned are already available in form of docks, which one might already have.

Yeah, a i5-6600k is a bit over $200 right now, and Ryzen is just around the corner. If a TB3 enclosure was, say, $100 (then you'd add an appropriate ATX PSU) and there were reliable, tested cables (and OSX drivers), I'd be all over this. EGPU has been around for a few years now and has yet to take off. Hopefully TB3 will remedy this, but I'm not holding my breath. I think a lot of the blame is on Intel, and hopefully with TB3 now they'll fully support eGPU.

I had to make the same decision in June, and I built a Windows desktop instead of using an external GPU.

However, I hope that I won't have to do that again in the future. I'd rather have just one computer for the sake of simplicity. Hopefully external GPUs will become mainstream in a few years.

Sure, but there are a few additional factors. First, $400 is the "lazy price". I had some credit card points to spend so I didn't want to waste too much time looking for deals. But if you wait for rebates, buy used, or get the less powerful non-Ti 1050, you could probably get close to $300. Second, I already had a very powerful CPU in my machine (as well as 16GB of RAM) and I didn't want to downgrade. Third, I move around constantly and prefer to keep my belongings to a bare minimum, so a compact eGPU box was far more compelling than even the smallest gaming PC case.

I've put a lot of thought and research into this already, and you basically summarized my thoughts on the matter. Building a gaming desktop is a better answer today.

But in a year, when you can buy a commodity Chinese-made TB3 enclosure for $200 that-- with a SINGLE CABLE-- powers your laptop, offers a bunch of extra USB3 ports, ethernet, audio jacks, and can also take a high-end GPU for gaming, well, that's a much more attractive proposition.

Along these lines, I was also curious about the quote

"the kind of CPU I could buy for cheap would be comically underpowered compared to the i7 4850HQ I already had in front of me"

Specifically: will the high-end i7 in a MBP run at Full Turbo indefinitely, or does it throttle down hard after a couple of minutes at full load to avoid overheating, like the one in my Dell Precision does?

I'm suspecting the latter, due to the basic physics of heat dissipation from a high-TDP chip. In this case, a sub-$200 Haswell i5 will run circles around it.

TB2 is akin to using one of those x16 slots that's wired to x4 (or worse, since if you're using a Gen3 x16 GPU, you'd be running at Gen2 x4, which would be an eightfold reduction).

It'll depend on your workload, but generally, it is going to be visibly slower than the "baseline" performance for any high-utilization workload.

> Is TB2's lower bandwidth a bottleneck that would keep you from using a really beefy eGPU for these purposes?

What limits you is the power input and cooling capabilities of your TB2 box. A 1080 has a 180W TDP (before O/C), so you need to feed it up to 180We and dissipate that much heat, a Titan X increases that to 250W. TFA went with a 1050 because it only has a 75W TDP, and that could be achieved with their enclosure (by swapping out the power brick, the "out of the box" version only handles 30W TDP).

After that, it's mostly a matter of feeding the GPU enough stuff to do, giving it a small number of expensive tasks will utilise the limited bandwidth better than the reverse, which is why the efficiency of eGPU (TB2, but also TB3) increases as you increase rendering quality and the like (compared to internal) e.g. you might get 70% of internal at low settings and 90% at high.

When using a single GPU, as is the case here, you will not be able to saturate a TB2 slot in any reasonable deep learning workload. For multiple GPUs, sure, PCIe performance matters, but for single GPUs it is not a bottleneck for any practical scenario I can think of.

Right now, I am working with large volumetric datasets (think 100+GB) and even then I am only seeing ~200MB/s peak transfers, which are well within the capabilities of TB2. In my experience, large datasets are bottlenecked by the hard drive, which is not a problem for modern rMBPs.

Edit: I am using a 2.5GB/s NVMe SSD, so the 200MB/s is the raw bandwidth used when mass-evaluating batches on the GPU. For training, I am seeing around 70-80MB/s sustained.

The performance penalty for paging in data from system RAM is already high enough that most games are highly optimized to keep the necessary data in GPU RAM. Paging in random bits of data is more latency-sensitive than bandwidth-constrained. I'm not surprised that the penalty is ~25%.

Using the eGPU to drive the internal panel means copying the frame buffer back to the internal GPU which is where the latency and performance penalty come from; I'm not sure how it is implemented, it might actually be copying twice (once to system RAM, then again to the discrete GPU RAM). Having more lanes over the TB connection would probably help here, though it might not eliminate the latency.

The 13" would benefit a lot more than TFA I suppose, since then we'd upgrade from Intel GPU to "serious" GPU, and not just dGPU to eGPU?

Thinking about it more, addition to my other comment re. power input: above 75W graphics card don't draw power from just the PCIe slot (as it's limited to, you guessed it, 75W), so you end up needing a "PC" PSU with 6-pin and/or 8-pin PCIe Power Connectors to provide additional power to the GPU (which you'll still have to dissipate).

Won't this be a matter of writing drivers?

That one looks a bit more legit than the CD120100A I got. Perhaps I will send it back and buy this one instead!

Eh, because he still had to hunt down a kinda sketchy China PSU from Amazon and hope that it could do the rates wattage, spent twice as much on the enclosure as the GPU, etc for a ~3.5x upgrade. I'm guessing that for most people, if they're considering this, the few hundred extra to get as much performance as you can out of it makes sense. Rigging up an ATX PSU isn't too hard, and one could put it all into a SFF desktop case if you care about it looking nice. Apparently TB3 enclosures are coming which also address all of these shortcomings except price.

Eh, the Wolfe looks like vaporware at this point. They cancelled their kickstarter and their explanation was vague and unconvincing. Their website hasn't been updated fully to reflect the cancelled kickstarter. I don't see a release timeline anywhere, and there's no mention of how their software works at all (but they do allude to a software key, even though they imply it's plug and play?).

The way you undo autocorrect in iOS is backspace once. The black suggestion tip will appear with your original spelling and tapping it will revert the change.

There have been so many times in my "IT guy" history I would have loved to use my laptop's screen as a monitor for another computer. It's kind of crazy to have to track down a monitor just to see if a headless desktop is booting properly so I can remotely access it from the laptop that's sitting right next to it.

And my current MacBook has a retina screen, compared to the 1440x900 screen on my desktop. I'd LOVE to use my desktop's power on my Macbook's screen.

Non-hacky Thunderbolt 3 eGPUs like the Razer Core do offer a loopback mode where the video feeds back to the laptops screen over the TB3 cable.

The extra traffic on the wire does incur a small performance penalty over an external monitor though.

A few Sony Vaio models had that feature circa 2011 (HDMI in to display on the 16" WXGA (?) 1080p screen). I was very close to getting one.

Interesting-- I was not aware of that!

That's the same bandwidth as TB2, which will add another 10% performance penalty versus a native 16-lane PCIe gen3 slot, for a total of 20% performance lost. Still completely workable.

< $100 might be tough.

You've got to have a decently made chassis to accommodate and cool a full-sized card and a meaty power supply for the latest and greatest GPUs. And that's before you add any docking functionality for extra ports.

The cheapest thing I've seen to date is the Alienware Amp which sells for around $150 but uses Alienware's proprietary port, which IIRC, is a bit slower than TB3.

If it works in Mac OS sure, although of course the performance won't be as good. The author seems to only be running windows for these tests

No, this cannot be done. In theory, it should be possible with xhyve in the future, but at this moment, no PCIe or TB passing is possible because of the lack of support on the hypervisor side on desktop operating systems.

Yeah, though I imagine it would have to be a pretty weak gaming PC due to cooling issues.

No, Appe deliberately blacklists eGPUs in macOS. It'll work out of the box only on Windows.

According to this guide, yes.

https://www.reddit.com/r/apple/comments/5l3y0a/egpu_my_exper...

My use case is admittedly a bit esoteric but: a) I need a MacBook for work, b) I only have room for a single computer in my life, c) I already have a pretty powerful computer with a TB2 port. Plus, it's nice to keep things relatively simple.

I have a mid-2015 MBP, and have it connected to a 4K monitor. Looks beautiful, but there's some obvious jerking going on from pushing so many pixels.

An eGPU would help solve this issue completely, and I could even go for dual 4K monitors (or more).

I would imagine you can, but nothing gaming-grade because you're doing it over USB 2.0 which is even worse on the Pi than it is on the PC.