RDNA RX 6600 and 6950 XT over OCuLink - eGPU for GPD Win Max 2 Ryzen 7840U

This is the first of a few test runs with different GPUs connected to a laptop through OCuLink and Thunderbolt 3/USB4. Nvidia, Intel ARC, and older AMD cards are in the queue.

The tests were made with GPD Win Max 2 - Ryzen 7840U at 24W TDP with 32GB LPDDR5X memory running at 7500 MTs. When connecting the graphics card through OCuLink it was just detected by the system while for Thunderbolt 3 it got detected after a few seconds - I got an AMD XConnect notification.

Each GPU was tested in two display configurations - one with the laptop's own display (using part of the connection bandwidth to send the frames back) and the second one with an external display connected via HDMI to the desktop GPU for optimal performance.

Benchmark results

FFXIV Endwalker benchmark

FFXIV Endwalker MMO benchmark with eGPU likely hits CPU limit as this game is quite CPU intensive. With Radeon 780M iGPU at 24W TDP, the laptop reaches a 7166 score. With an external GPU, we instantly hit more than double. The Thunderbolt scores are much lower. I'll have to look more into this. This could be a ReBAR issue or something else - like an actual bandwidth/compatibility issue for RX 6950 XT.

3DMark Time Spy benchmark
3DMark Night Raid benchmark

3DMark Time Spy results are kind of what would be expected. 780M gets 3093 while RX 6600 is more than double that. RX 6950 XT is more than double of RX 6600. CPU scores are consistent across configurations. For Thunderbolt are bit lower due to the CPU having to handle it combined with other I/O.

3DMark Night Raid results are quite different. When the laptop's own (internal) screen is used the GPU results are significantly lower. Either this benchmark uses all the PCIe throughput or there is a bug or incompatibility.

3DMark Solar Bay

3DMark Solar Bay is a cross-platform lightweight ray tracing benchmark. DXR can be more PCIe bandwidth-intensive and it seems to be visible here where both internal configurations reach much lower scores.

Unigine Valley benchmark

Unigine Valley is an old DX11 benchmark. Internal OCuLink configurations reach only half the FPS of external ones. Thunderbolt 3 seems to be impacted really hard.

Unigine Superposition

Unigine Superposition is a newer DX12 benchmark and it seems to handle internal configurations really well. Thunderbolt has some severe problems, where RX 6950 XT reaches less than internal 780M can.

3DMark Wild Life

3DMark Wild Life is a cross-platform Vulkan based benchmark. For this one internal screen also got hit really hard.

PCIe bandwidth test

3DMark PCIe bandwidth test tries to saturate the PCIe bandwidth available and get an accurate measure of an actual limit. It ran fine for RX 6600 but for RX 6950 XT it had problems and ran really slowly. This could be related to the OCuLink cable, adapter, or similar issue. Some users do have problems reaching max throughput with some OCuLink cables and on GPD Discord there is quite a bit of discussion about this.

API overhead benchmark

3DMark API overhead test is and old benchmark that was supposed to show how DX12/Vulkan perform vs DX11 on given system. As DX12 and Vulkan are very performant they can be bottlenecked by short PCIe connections and thus measure such connections. Here the RX 6950 XT over OCuLink does show the same problem as in the PCIe bandwidth test.

More testing and troubleshooting

Those are just some quick results from the eGPU setups. They did show a few issues to look into like the connection speed. Other content creators also pointed out ReBAR On for Thunderbolt 3 is decreasing the performance so it's also a thing to check. Then some actual games to see if there is a pattern on which games performing good/bad on internal screens or bad overall for eGPU.

I also plan to test Nvidia and Intel ARC cards as well as other AMD cards - Vega and Polaris. If you have any suggestions on what to use and test feel free to comment or contact me.

Comment article