Testing several high end GPUs in eGPU in 2023 - a Report

As I could barely find any up to date information about using eGPUs with 30XX and 40XX Nvidia GPUs online I set out to test and check for myself.
Some eGPU enclosure manufacturers maintain compatibility lists but I found them to be incomplete/outdated or simply inaccurate. The zenith of using eGPU seems to have been a around 4 years ago and now nobody really cares about them any more, am I wrong? I wanted to find out if they actually still work with todays GPUs and how their performance compares to the Desktop or mobile edition GPUs with the same name.

I acquired an AKiTiO Node external GPU enclosure and then tried several older and newer GPUs with different laptops.

First I tried RTX 2080 (Blender benchmark points):

  • Desktop PC with RTX 2080: 2350
  • Laptop with built in RTX 2080 mobile: 1570
  • Laptop: XMG neo e20 with eGPU RTX 2080: 2238
  • Laptop: Dell XPS with eGPU RTX 2080: 2205
  • Laptop: Dell 7400 with eGPU RTX 2080: 1840

The conclusion here was that rendering on eGPU with RTX 2080 was almost identical to the Desktop performance and faster than the 2080 mobile GPU even on rather low end hardware (Dell 7400). The XMG neo and Dell XPS are rather high end machines and have mobile edition NVIDA GPUs built in. Note though that using the Dell 7400 with attached eGPU was no fun as everything was sluggish and slow, even just clicking and dragging windows around took several seconds.

Next up, RTX 3090 (Blender benchmark points):

  • Desktop PC with RTX 3090: 6372
  • Dell XPS with eGPU RTX 3090: 4979

As an RTX 3090 mobile GPU does not seem to exist I have no comparison values here. The conclusion is that the performance is much worse on the eGPU compared to the Desktop GPU (just around 78%) compared to much better ratio of around 94% on the RTX 2080.

Next up, RTX 4090 (Blender benchmark points):

  • Desktop PC with RTX 4090: 13086
  • Dell XPS with eGPU RTX 4090: 10072
  • Dell 7400 with eGPU RTX 4090: 8475
  • Official benchmark results from database for RTX 4090 mobile: 8143

In conclusion the more recent a GPU seems to be the worse it performs in an eGPU, the RTX4090 only reaches around 77% of the desktop performance in an eGPU, for the RTX 3090 it was around 78% and for the 2080 even ~94%.

The RTX 4090 was a bit of an adventure in the eGPU enclosure. Luckily this particular model fit the overall length of the eGPU enclosure but there was a piece of metal near the output ports that prevented the card from being inserted all the way. I had to take the metal piece off which prevented the card from being held i place with screws. Also the RTX 4090 has a different/new power connector and with an adapter cable requires 3x PCIe power inputs, the eGPU enclosure only supplies two and as tested 2 was indeed not enough to spin up the card. So an additional external ATX power supply was added to the mix and miraculously the RTX 4090 booted up in the eGPU enclosure and was properly recognized by the connected laptop. Not an actually practical solution for a useable setup though…

Overall my impression was that laptops that already have an Nvidia graphics card installed seem to be able to handle the eGPU better (NVIDIA Optimus?) than a laptop that only has Intel GPU (Dell 7400 for example was very sluggish in general operation - is practically unusable).

I hope this helps someone who then does not have to go through the lengths of trying this all by him/herself like I did :slight_smile:

4 Likes

Hmm… but weirdly enougth you does not mention anything about how the eGPU is connectiong the via PCIe lanes… (only power connectors…)

The PCI Express standard defines link widths of x1, x2, x4, x8, and x16… so of course you can connect a high performance graphics card with x1… and it will show you something… but the newly updated data feed by the CPU will… take a bit longer…

Ah yes, a good point.
The Akitio Node uses a Thunderbolt 3 connection with 4 lane PCI Express 3.0

High end GPUs want 16 lanes PCIe 4.0 so its expected that there is a connection/bandwidth bottleneck, what I wanted to find out was how much this affects rendering in contrast to gaming where transfer time is a major issue, for rendering I figured that once a scene is in the VRAM the bandwidth has less of a performance impact.

Did you also test the Desktop PC using the eGPU, as well as that same Desktop PC with the GPU internally installed? That would tell you very clearly how much you lose to bottleneck, versus how much additional you lose to the switch to the rest of the laptop hardware.

1 Like

Great tests and surprisingly good results.

1 Like

That would indeed have been an ideal test but unfortunately one additional lesson learned was that while every modern PC/Laptop has USB-C connectors not all of them actually have a Thunderbolt port… :grimacing: