A common issue for laptop users is the lack of GPU power. Even the fastest mobile GPUs, in SLI or Crossfire cannot reach the echelons of performance of a higher-end desktop, mainly due to the power consumption and heat generation.  Not only that, laptops with high-end mobile GPUs in order to cope with heat generation tend to be far from portable. Sure, they are still easy to carry around compared to a full-size desktop system, but not many are willing to carry one around on a daily basis. In other words, if you want a laptop that's relatively portable, you are left with mediocre GPU performance that usually doesn't satisfy the needs if you happen to be an active gamer. 

Ever since the original Thunderbolt was released back in 2011, there has been a lot of discussion about the potential of using Thunderbolt for external GPUs. Today's mobile CPUs are far more than capable of driving desktop GPUs and as Thunderbolt is essentially just PCIe and DisplayPort in a single interface, a laptop with an external GPUs makes almost too much sense. 

SilverStone's/ASUS' Thunderbolt eGPU enclosure at CES

So far a handful of companies, such as MSI and SilverStone, have showcased their external Thunderbolt GPU enclosures at trade shows, but due to issues such as performance and hot-plug applications, no-one has made it to retail . Intel's decision to double the bandwidth with Thunderbolt 2 negated the launch of the original Thunderbolt-based designs, although with any luck TB2 should be an appropriate drop-in. Especially with GPUs, bandwidth can make a dramatic difference in performance and given the niche of external Thunderbolt GPUs, many users wouldn't have been satisfied with a product that doesn't provide at least near the maximum performance. 

Another big issue is obviously driver and operating system support. To make matters worse, nearly all Thunderbolt-equipped devices are Macs and traditionally Apple likes to have very tight control over drivers and other elements of the OS, making it hard (or even impossible) to develop an external GPU that would also function under OS X. In the PC arena, a few motherboards and products exhibit Thunderbolt support, and it is primarily up to Intel working with Microsoft to develop Windows based drivers.

DIY to the Rescue!

Disclaimer: All information and results here are based on a forum post and a (now private) Youtube video. We cannot guarantee that the results are accurate, thus any and all purchase decisions must be done at own risk with the possibility that the results may or may not be on par with what is reported below. 

As no company has set forth and commercialized their products yet, the enthusiast group has been looking for a do-it-yourself method to drive an external GPU over Thunderbolt. I came by a very interesting setup over at Tech Inferno forums today and thought I would share it with a larger readership here. A forum member squinks has managed to run an NVIDIA GTX 780 Ti over Thunderbolt 2 using Sonnet's Echo Express III-D chassis with Corsair's RM450 power supply dedicated to the GPU. 

Courtesy of Tech Inferno forum user squinks

Update: The video of the setup in action originally went private right before our post but it has now been made public again and can be seen here.

The results are certainly auspicious. Based on squinks' own tests and GTX 780 Ti reviews posted online, the performance seems to be around 80-90% of the full desktop performance based on synthetic benchmarks (3DMark and Unigine Heaven). Given that Thunderbolt 2 offers only 20Gbit/s of bandwidth while a PCIe 3.0 x16 slot offers 128Gbit/s, getting 80-90% of the performance is a lot more than expected. This will vary depending on the game, as based on our own PCIe scaling tests the PCIe bandwidth may cause little to no difference in some games while in others the drop can be close to 50%. The more severe the needs of the PCIe connection, the worse the performance.  Either way, Thunderbolt 2 can show a potential for external GPUs even when talking about the most powerful ones that should also require the most bandwidth. 

According to the forum posts, the setup is also pretty much plug and play, but only as long as the GPU is connected to an external monitor. Once everything has been connected and drivers installed, the GTX 780 Ti will be recognized as a GPU like it would in any desktop system. Getting the external GPU to drive the internal display is also possible, although there apparently seems to be some limitations with this homebrew method. First off, it only works if the computer doesn't have a discrete GPU as then NVIDIA Optimus can be used to enable GPU switching. If there is already a discrete GPU in place (like the GT 750M in the high-end 2013 rMBP), then Optimus cannot be used and unfortunately you'll be limited to an external monitor. Secondly, there seems to be some loss in performance (~5-20% in addition to the original loss from Thunderbolt 2) when driving the internal display, which is likely due to Optimus and its limitations. 

The big question is whether such setup reasonably affordable in any way. Currently, the short answer is no. The Sonnet Echo Express III-D chassis alone costs $979, and you'll need to add the cost of the GPU and power supply to that. The weight of the chassis is also 7.5lb (3.4kg), without the GPU or the power supply, hardly making it all that portable.  In total this means ~$1500 minimum if you are going with a higher-end GPU (which you should given the cost of the chassis). For comparison's sake, I quickly gathered parts for a decent gaming rig in NewEgg and the total came in at $764.94 (without GPU and PSU). That's with a Core i7-4770K, ASUS Z87 motherboard, 8GB of DDR3-1600, 120GB SSD, 1TB hard drive and mid-price case, so we are not even dealing with a budget system. In other words, you can build a higher performance system for over $200 less and take full advantage of your GPU. 

Update 2: As some of you mentioned in the comments, there are cheaper alternatives available that provide about 70-90% of the desktop performance. What you need is a Thunderbolt to ExpressCard adapter (like Sonnet's Echo ExpressCard Pro) and an ExpressCard to PCIe adapter (like the BPlus PE4L V2.1), which together come in at $240 when bought straight from the manufacturers' online stores. Add a cheap ~400W power supply to that and the total is less than $300 (without the GPU, of course). If you are interested in external Thunderbolt GPUs, I recommend that you take a good look at Tech Inferno forums as they have several guides and other resources from troubleshooting to benchmarks.

All in all, it is fun to see an external GPU connected via Thunderbolt 2 can actually get up and running. The price and DIY-ness are currently factors that don't exactly allure the masses, but there is a potential market for a retail product that is designed specifically for this. Pricing is, of course, a major factor and at $200-$300 I could see external GPUs gaining popularity but once you go over $500 it, in most cases, becomes more viable to build a dedicated gaming rig. 

Source: Tech Inferno forums

Comments Locked


View All Comments

  • Cygni - Sunday, May 4, 2014 - link

    Really hope this comes down in price a bit more. Pretty much turns a laptop with a solid CPU + Thunderbolt 2 connector (13in Retina?) into the ultimate computing situation. Use it on the couch or on the road, and its cool, portable, silent, comfortable. Get home, plug it into the base station and a big monitor, and its a powerful gaming machine too.

    Very cool.
  • Nexing - Monday, September 22, 2014 - link

    Cannot comprehend why worrying laptop Manufacturers have not grasped such rather novel usage for an perfectly attainable tech since a while, and tried to sell it already...
  • kmmatney - Friday, March 6, 2015 - link

    It seems easier just to buy a laptop with a secondary gpu already in it? Then you have it when you travel. The price difference is probably less that what you have to pay for this. I have an NVidia 650M in my work laptop, and it games reasonably well, and is with me in my hotel room when I travel.
  • Khenglish - Monday, May 5, 2014 - link

    Kristian you should take a look at the expresscard/mPCI-E results (the majority of results on TechInferno). The adapter (PE4L 2.1b) only costs $70, and Nvidia drivers switch to a compressed PCI-E mode when detecting an x1 link, roughly doubling performance up to thunderbolt levels despite half the bandwidth. Typically synthetics on an x1 2.0 link are slightly faster than thunderbolt, but games are usually slightly slower.
  • Calista - Friday, May 9, 2014 - link

    It's still a quite limited bandwidth and so it will cripple most high-end cards. So we're left with an external solution that needs a high-end card to make sense (since integrated graphics is getting better by the day) while at the same time the better the external card the more starved on bandwidth it will be relatively speaking.
  • Tegeril - Monday, May 12, 2014 - link

    There is a colossally large chasm between integrated graphics and high end cards right now. $150-200 cards run circles around integrated.
  • dflippo - Monday, May 4, 2015 - link

    No, the most compelling reason for this over integrated graphics is: HEAT.

    Most normal laptops can't handle CPU + GPU heat for gaming. Offloading the graphics load to a dedicated card with it's own thermal ecosystem is perfect for laptop gaming. Maybe not in an airport or on an airplane, but at home, or in the hotel, it would be great.
  • Impulses - Monday, May 5, 2014 - link

    The logistics aren't the issue hereDIYe fact that we have to resort to DIY at all is... Intel should be all over this kinda thing, pushing for it to happen, not like they have a discrete GPU business to cannibalize. If anything this would sell a lot more Intel laptops if it was easier to setup and mainstream.
  • SirKnobsworth - Monday, May 5, 2014 - link

    Some of the performance hit may come from the fact that the ThunderBolt PCIe lanes are usually connected to the chipset and not directly to the CPU the way most GPUs are. Not that this is going to change any time soon.
  • repoman27 - Tuesday, May 6, 2014 - link

    Actually, just the opposite is true. The lion's share of PC's with Thunderbolt ports are Macs, and most Macs have the Thunderbolt controller connected to PCIe lanes coming from the CPU, not the PCH. I think iMacs with discrete graphics are the only examples of Apple using the PCH lanes instead.

    No matter how you slice it, 1.385 GB/s (the real world PCIe throughput limit of Thunderbolt 2) is less than ~13 GB/s (the maximum throughput of a typical PCIe x16 slot). However, PCIe traffic is often bursty, and like memory bandwidth, with decent caching, after a certain point you see seriously diminishing real world performance benefits from further increases in bandwidth alone. Latency is certainly a factor, though, and Thunderbolt is pretty brutal in that regard—around 1.5 µs per hop versus a few hundred ns for a typical PCIe switch. Obviously you would want a Thunderbolt connected GPU to be the first device in a chain, or else you could be facing as much as 9 µs of round-trip latency for the last device in a 6-device chain. If GPU drivers were in any way optimized for lower bandwidth / higher latency connections like Thunderbolt, I'm sure the performance gap could be narrowed even further though, despite having 1/10th the bandwidth and 5x the latency.

Log in

Don't have an account? Sign up now