Est. eGPU Performance based on PCIe Bandwidth Scaling

The following estimates (~) are based on this article on PCIe bandwidth scaling for current high-end gen GPUs. I am extrapolating data specifically for the GTX 680.

Note: I am not entirely sure multiple Thunderbolt ports could be bonded together.

Interface	Bandwidth	GPU Performance	Notes
ExpressCard 2.5Gbps	0.31GBps	~63%
ExpressCard 5Gbps	0.63GBps	~67%
PCIe 1.1 x4	1GBps	73%
Thunderbolt 10Gbps	1.2GBps	~76%	Haswell/Redwood Ridge(2013)
PCIe 2.0 x4	2GBps	88%
PCIe 1.1 x8	2GBps	88%
Thunderbolt 10Gbps x2	2.5GBps	~92%	Haswell/Redwood Ridge(2013)
Thunderbolt 20Gbps	2.5GBps	~92%	Broadwell?/Falcon Ridge(2014)
PCIe 1.1 x16	4GBps	95%
PCIe 3.0 x4	4GBps	96%
PCIe 2.0 x8	4GBps	96%
Thunderbolt 20Gbps x2	5GBps	~97%	Broadwell?/Falcon Ridge(2014)
PCIe 3.0 x8	8GBps	100%
PCIe 2.0 x16	8GBps	100%
PCIe 3.0 x16	16GBps	101%

<style>#newspaper-a{font-family:"Lucida Sans Unicode", "Lucida Grande", Sans-Serif;font-size:12px;text-align:left;border-collapse:collapse;border:1px solid #69c;margin:20px;}#newspaper-a th{font-weight:normal;font-size:14px;color:#039;border-bottom:1px dashed #69c;padding:12px 17px;}#newspaper-a td{color:#669;padding:7px 17px;}#newspaper-a tbody tr:hover td{color:#339;background:#d0dafd;}</style>

Expect results to vary depending on application and device.

A lower performing card should see less variation between interfaces and skew closer to 100% on each interface.

A higher performing card will show more variation between interfaces and skew away from 100% for each drop in bandwidth.

 

sorry but that's bs. gpu performance is heavily based on the aplication you are using as well as on the graphic card. for the same interface you can have 25% of variation.

 

Agree, every GPU has a specific bandwidth that some Interfaces can handle and other that don't, changing the performance lost
And no, there's no way to combine 2 thunderbolts now to create a 2.0 x8 interface
Although I've encountered few articles saying that you can use 2 thunderbolts to create an SLI, but both interfaces need to have at least 2.0 x4 for it to work...
I'm now creating a TH05 thunderbolt GTX 680 FTW 4gb 2.0 x2 now... when they release one with 2.0 x4 I'll try SLI and say on forum

 

even thunderbolt x8 will still bottleneck some aplications and gpu's. a x16 version would probably be 100% bottleneck free.
it's a very slow interface when compared to pci-e and external pci-e. still, the fastest external interface we have on notebooks, unfortunately.

 

Yeah, GTX 690 have a bottleneck with loss of 1% on 2.0 x16 now compared to 3.0 16x
Next gen GTX due to launch in march will probably have a bigger gap

 

If you'd read the article I linked to, you would see that they tested multiple applications and they did get different results.

But, for the most part, the variations in performance were pretty consistent based on the bandwidth of the interface.

The data that I present is extrapolated from the result aggregates (of the GTX 680).

That's not to say that performance isn't going to change depending on the application you are using, but that, for most applications, these should be fairly accurate.

As for variations between different graphics cards, I thought it was implicit.

Of course a lower performance card will be less affected by restricted bandwidth. If you had read the article you would have seen this is specifically addressing the GTX 680.

I clarified this in my initial post.

 

The results I give do show a bottleneck.

I think you are confused by my "x2" notation. That was referring to bonding two separate Thunderbolt connections not that it is running on a PCIe x2 bus.

Thunderbolt is run on a PCIe 3.0 x4 bus, but is actually limited to a little over x1 per port. I've showed the the equivalent bandwidth in the chart.

I know it's a slow interface in comparison to pcie, but you must also keep in mind that even current high-end gaming is not saturating x16 PCIe.

Therefore there are going to be diminishing returns as you get into higher bandwidth performance and a 20Gbps Thunderbolt connection should be able to offer enough bandwidth for most applications.

That's not to say that something like bitcoin mining (or another GPGPU application) might not show a larger gap in performance. Extremely bandwidth intensive applications will of course have larger bottlenecks.

These are simply estimates based on aggregate performance (specifically for games).

 

I should have made it clear that these results are only applicable to the GTX 680.

The article makes it clear what GPU they are testing.

Obviously for an older/lower performance card the eGPU performance will show less variation and skew closer to 100%.

For something like the GTX 690 I'd expect them to skew the other direction.

But I thought all of that was obvious, so i chose not to elaborate. I have clarified my post.

 

it's not related to the card performance but rather on the bandwith. so a cheaper card might even perform better then a more expensive one. it's only related to bandwith. For example 2gpu cards perform very badly because they need much more bandwith and are heavily bottlenecked.
funny is, external pci-e spec has been around since 2009 at least with the same performance of pci-e and yet no one implements it on notebooks.

 

I find it rather sad, not funny. With more and more OpenCL applications surfacing, the manufacturers are making sure they milk the last dollar out of us by making us buy complete systems instead of GPU upgrades. Commodore went the route of catering to the wishes of the consumers, on the Amiga 500 they had an expansion interface called "Zorro" which allowed expansion cards to be connected to the computer externally. You know how they ended up...

 

there's nothing wrong with a company giving customers what they want. I'm 200% sure that when thunderbolt becomes a mainstream interface just like usb (and it will only because intel wants it and they are the ones who set the rules (though they don't really care anything about egpu's)) that the first company with a moderate price (sub $50 (because it costs about $20 to manufacture)) thunderbolt egpu will make tons of money.
Also if a company started selling notebooks with external pci-e interfaces it would sell like hot cakes. you could have a small sub 1kg 11" notebook with the same performance of an high-end desktop.or workstation. The market is huge. There's lot's of money to be made.

 

I too wish for ThunderBolt or external PCI-E to become prevalent...