2021 15" ASUS TUF Dash F15 [11th,4C,H] + RX 6900 XT @ 32Gbps-M.2 (ADT-Link R43SG) + Win10 2004 // my 1st RX 6900 XT build
My previous build with this ASUS TUF Dash F15 + AORUS RTX 3080 Gaming Box posted very good Thunderbolt eGPU performance [build link]. The on-die Thunderbolt controller on Tiger Lake H-CPU provides a dedicated and direct-to-CPU connection for the eGPU. I wanted to see how these improvements through Thunderbolt 4 would compare to the M.2 eGPU interface. My configuration has the i7-11370H quad-core 35W TDP, RTX 3070, 16GB RAM, and 1TB NVMe. I'd recommend the lower configuration with RTX 3060 dGPU for $350 less [Best Buy].
2021 15" ASUS TUF Dash F15 – i7-11370H/RTX 3070 dGPU/Iris Xe Graphics iGPU/16GB RAM/1TB SSD
I was hoping for PCIe 4.0 M.2 slots in the ASUS TUF Dash F15. Unfortunately, that PCIe 4.0 x4 connection is reserved for the RTX 3070 dGPU. The two M.2 slots are through the PCH and run at PCIe 3.0 x4. The plastic bottom cover with standard Philips screws make closing and opening this laptop a quick task.
You may see the Dell DA-2 220W power adapter in the photos above. However, that was my wishful thinking. As soon as I ran GPGPU test in AIDA64, the RX 6900 XT crashed immediately. I swapped it out for a 450W ATX PSU. This power supply went through several tests but was no match for the RX 6900 XT either. I retrieved the FSP 750W ATX PSU out of the Mantiz Saturn Pro which finally handled this GPU with ease.
I ran all benchmarks off an external SSD, Samsung T5 500GB. Both RTX 3070 dGPU and RX 6900 XT eGPU connected to the 43" Frame TV for monitor output (4K @ 60 Hz over HDMI). The laptop internal display is FHD so in order to test loopback at higher resolutions, I connected the 4K TV to the laptop's HDMI port. AMD XConnect worked well to accelerate the TV through eGPU. The computer performance level was set to Turbo in Armoury Crate.
|RTX 3070 dGPU||RX 6900 XT M.2 eGPU Loopback||RX 6900 XT M.2 eGPU External||RX 6900 XT TB3 eGPU External
The RX 6900 XT performance is incredible through M.2 eGPU adapter. I have started running this same graphics card through the Mantiz Saturn Pro for Thunderbolt 3 eGPU but the performance is less impressive partly due to Radeon drivers. This is an issue I've noticed since spring last year. Anything newer than Radeon driver version 20.4.2 is inconsistent through Thunderbolt eGPU mode.
Looking closer at the numbers, you can also see the RX 6900 XT eGPU improvement over RTX 3070 dGPU is much better in 3DMark vs Unigine. I will be adding more results through the Thunderbolt 3 enclosure, Mantiz Saturn Pro. Due to tariffs and limited supply, this graphics card costs more than the laptop. It's a ridiculous situation and I hope it's over soon.
Context: I am moving overseas, will be inbetween AirBnBs for a while and don't want to take both my Node and a ADT-Link + PSU
EDIT: I realize the TB performance won't be equal as it is connected through PCH on my machine.
@jake_laurie, 20.4.2 doesn't have support for RX 6000 series cards so we have to use newer versions. I will be posting more TB3 eGPU results with the same drivers version. Thunderbolt eGPU is much more convenient with hot-plug, Power Delivery, and expansion ports so it's better for your trip overseas.
I added Thunderbolt 3 results in the OP. The RX 6900 XT is too large to fit inside the Mantiz Saturn Pro. It was running without the outer enclosure. Unigine numbers are all over. Superposition for example stayed at exactly the same score at different resolutions. I even disabled iGPU & dGPU so that only eGPU was active to make sure Windows wasn't using the wrong graphics card.
@itsage , some big increases in the CPU/Physics numbers of the added 3dmark TB3 results compared to the previous set. Wonder what changed on the system and is this also boosting the graphics numbers?
@nando4, It was the same drivers as M.2 testing. Amoury Crate was in Turbo mode as well. The only difference is a fully closed bottom cover compared to a small gap on the side of M.2 adapter cable.
It would be interesting to see a Turing/Ampere card like the 2080Ti or 3080 in these two implementations and see how those compare vs the RDNA2 card. That Superposition score with the TB connected 6900XT seems so low that it can't be accurate.
My system scores better than that in that test and my i7-8565U is currently thermal throttling while undervolted at 10W in Cinebench. Waiting on delivery of some TIM so I can do a refresh TIM application. It sounds like the wildly inconsistent performance might be a driver issue, strange. I would expect the Dash using on-die TB4 + 6900XT to be far better than my inefficient 8th Gen + 2070 system. Are you sure it wasn't using the Xe iGPU instead of TB?
Also it looks like the Dash w/ 3070 80W dGPU beats my setup in all three tests you ran, 20.3% at 1080p, 13% at 1440p and 7.6% at 4K. I'll run tests again after I get a chance to reapply TIM to my CPU and get it back up to the 15-16W it's supposed to be drawing. LOL.
@odin, It's a well documented issue with Radeon driver newer than 20.4.2. The eGPU performance is very inconsistent. I ran Thunderbolt 3 tests on the RX 3080 Gaming Box with the ASUS TUF Dash build link]. The liquid-cooled RTX 3080 in this Gaming Box is not easy to be extracted from the enclosure to run in M.2 interface.