Gigabyte Thunderbolt 3 - RX 580 Gaming Box

#5529500 Posted on: 03/19/2018 11:53 PM

---

I think what you said is heavily dependent on whether you're plugging a display directly into the eGPU or trying to use your laptop's display. If you are expecting to use the laptop's built-in display (especially if it is 1080p or higher) with high-res textures, to my understanding, that soaks up a lot of your usable bandwidth, where a 580 would be bottlenecked.

Keep in mind that in a lot of cases, a GTX 1080 (non-Ti) starts to see performance losses when going from x8 PCIe 3.0 lanes down to x4 lanes at various resolutions, when using a display connected straight to the GPU. PCIe 3.0 with x4 lanes is roughly 55Gbps. If we were to dramatically generalize that a RX 580 uses proportionately less bandwidth than a 1080 (when accounting for its lesser performance), I don't think it is possible to consistently take advantage of it's processing power over USB 3.2, regardless of which display is used. But, I could be wrong.
While true, that bottleneck may still be very acceptable as long as graphics does not need to load bunch of textures from system memory every frame. For years, people use PCIe 2.0 and 3.0 x1 in notebooks (meant for Wifi and such) for external PCIe GPUs. Ugly, but relatively cheap solution. And it somewhat works.

I tried to check actual IO utilization, but HWiNFO64 always shows this value as 0%. So lovely configuration for graph displayed over time and nothing to be shown :/
Maybe someone with nV card or older driver/OS can try, as it may be some bug with insider build I have.

#5529510 Posted on: 03/20/2018 12:27 AM

---

True, but it's the bottleneck why I specifically feel a 1060 or 570 make for a better option - you pay less without really any performance loss in most cases. Meanwhile, weaker GPUs use less power. This is important to factor in, since the power bricks get disproportionately more expensive, large, and hot as you increase the wattage. From what I recall, the price really starts to go up fast once you breach 80W. If you go for an RX 550 or GTX 1050Ti, those are efficient enough to not need any PCIe power connectors, (therefore should be sub-75W) and a power brick for that should be relatively cheap. They're also slow enough that as long as you lower texture details, they should perform well on USB 3.2. They still might be a little too much for PCIe 3.0 x1 slots, though.

I'm not sure what your level of interest in all of this is, but there is something I could recommend to you, that maybe isn't elegant but very cheap. There are products on eBay that allow you to convert M.2 slots into x4 PCIe slots. I think they cost something like $5USD. You can also get M.2 riser cables. Combine these together and buy some 12v power brick and you could create your own eGPU solution for laptops for about 1/10 the price of these other eGPU enclosures. Since you're feeding directly from PCIe, there is no latency loss. The downsides are the ugly appearance, fragility, and not hot-swappable.


What were you testing? If you're not really doing anything, PCIe usage tends to stay from 0-1%.
BTW here's what I was referencing earlier:
https://www.techpowerup.com/reviews/NVIDIA/GeForce-GTX-1080-PCI-Express-Scaling/
They seem to do one of these tests every other year. Pretty interesting and comprehensive - I don't know of anyone else who does tests like this.

#5529519 Posted on: 03/20/2018 12:52 AM

---

Don't waste your money. If this is your only option, you need to rethink your life decisions.
Harsh but true.

#5529520 Posted on: 03/20/2018 12:54 AM

---

True, but it's the bottleneck why I specifically feel a 1060 or 570 make for a better option - you pay less without really any performance loss in most cases. Meanwhile, weaker GPUs use less power. This is important to factor in, since the power bricks get disproportionately more expensive, large, and hot as you increase the wattage. From what I recall, the price really starts to go up fast once you breach 80W. If you go for an RX 550 or GTX 1050Ti, those are efficient enough to not need any PCIe power connectors, (therefore should be sub-75W) and a power brick for that should be relatively cheap. They're also slow enough that as long as you lower texture details, they should perform well on USB 3.2. They still might be a little too much for PCIe 3.0 x1 slots, though.

I'm not sure what your level of interest in all of this is, but there is something I could recommend to you, that maybe isn't elegant but very cheap. There are products on eBay that allow you to convert M.2 slots into x4 PCIe slots. I think they cost something like $5USD. You can also get M.2 riser cables. Combine these together and buy some 12v power brick and you could create your own eGPU solution for laptops for about 1/10 the price of these other eGPU enclosures. Since you're feeding directly from PCIe, there is no latency loss. The downsides are the ugly appearance, fragility, and not hot-swappable.


What were you testing? If you're not really doing anything, PCIe usage tends to stay from 0-1%.
BTW here's what I was referencing earlier:
https://www.techpowerup.com/reviews/NVIDIA/GeForce-GTX-1080-PCI-Express-Scaling/
They seem to do one of these tests every other year. Pretty interesting and comprehensive - I don't know of anyone else who does tests like this.
It's just a generic technology interest. I saw those PCIe 3.0 x1 cables and hacks people did to have it working. M.2 slot to PCIe 3.0 x4 would in most cases have no issue. Except I think that native PCIe in both cases have trouble with turning card ON/OFF. Hotswap behavior is unusual for PCIe in consumer segment.

Anyway. In test you linked, there comes following info:
PCIe 1.1 x4 (1GB/s = 8Gbps) loses around 33% of GTX 1080 performance in worst case scenario. Best Case Doom 19% loss.
PCIe 2.0 x4 (2GB/s = 16Gbps) loses around 15% of GTX 1080 performance in worst case scenario. Best Case TR 5% loss.
(And those are on 1080p gaming, where mobile CPU is more likely to be bottleneck "instead".)

So I would say that USB 3.2 is quite acceptable with 20Gbps.

#5529533 Posted on: 03/20/2018 01:28 AM

---

I've done a few experiments myself with x1 slots and haven't had any trouble with them (in terms of them being on or off). But, I've only tested such situations in Linux. I intend to try some tests with one of those M.2 to x4 converters but I'm not sure when I intend to get around to it. Like you, I'm more interested in the technology than actually having a need for it.


Don't forget though - there's also the bandwidth of transferring the video signal back to the laptop, if you intend to. That 5% loss in the best-case-scenario will increase significantly. But as stated before, if you intend to use a separate monitor then yes, USB 3.2 ought to suffice.

Another thing to consider that I forgot about is there's some overhead due to the USB translation layer, which will result in some performance loss. This is why, for example, there aren't any external USB 3.1 drives that can meet or exceed SATA III, despite the 4Gbps lead. To my understanding, Thunderbolt is a lower-level connection than USB and doesn't suffer the same overhead.

Something I'm not too sure about is how bandwidth is allocated by the device. So more specifically, if you have a PCIe 2.0 GPU, to my understanding, the GPU is getting a more bandwidth out of a 2.0 x16 slot than a 3.0 x8 slot, even though the total bandwidth of the slots are roughly the same. From benchmarks I've seen, it seems Thunderbolt will actually split up the bandwidth into multiple lanes, but I'm not sure if a USB adapter will. If it can, that'd be awesome.

Anyway, interesting discussion. I personally find this stuff fun to think about - I hope I'm not coming across as annoying or arrogant, but I find you're good to bounce ideas off of.