News External GPU Dock Wreaks Havoc on RTX 4090 Gaming Performance

[Admin]

YouTuber discovers incredibly poor performance from Nvidia's RTX 4090 on an external Thunderbolt 3 GPU dock. With the GPU dock and the laptops tested, the 4090 wouldn't push past 60% of its actual performance potential.

External GPU Dock Wreaks Havoc on RTX 4090 Gaming Performance : Read more

 

[ManDaddio]

I did not watch Jared's video. I'm just coming out this with my own thoughts on the subject.
Why would anyone even consider using today's high-end GPUs with a laptop? An RTX 2080 TI, 6800xt,or RTX 3080 should be the maximum performance cards used. And even the latter two would be overkill. Yes I understand people use all kinds of cards with the docking. I'm just coming from a practical stand.
Also for the price that you pay for the docking station and the card people should just go out and buy a high-end laptop. That would be more practical and efficient.
I can't say much for Apple users because their choice is very limited. Maybe for them this solution is a solution. But I still wouldn't put today's top tier cards in a docking station for everyday users and gamers.

 

[Llamaking2]

The article states:

Unfortunately, Jarrod neglected to share details on the external GPU dock he is using, only noting that it's a Thunderbolt 3 dock -- which is good considering most GPU docks today are using Thunderbolt 3.

Although in the video embedded states pretty clearly at 0:14 it is a Razor Core X Chroma dock that he had to modify slightly to get the 4090 to fit inside of.

 

[tek-check]

Did anyone expect anything else??? it is PCIe 3.0 x 4 vs PCIe 4.0 x 16. That is an eight times transfer rate difference. Thunderbolt 5 with PCIe 4.0 might stand a little bit better chance with double the transfer rate of Thunderbolt 3.

Exactly. In addition, 4090 would perform better in desktop Gen3 x4 slot as Thunderbolt also has an overhead.

 

[NTR]

Unfortunately, Aaron Klotz can't even extrapolate information 20s into the video YIKES...

 

[escksu]

Not surprising. I would say it's definitely has to do with the resizable bar and thunderbolt bandwidth. Since tb 3.0 is essentially a pcie 3.0 4x slot, 4090 is definitely going to struggle on it, esp. if there is resizable bar enabled. TB has its own overhead so actually bandwidth is even lower than that of a pcie slot.

 

[Keng Yuan]

Asus should make their proprietary eGPU port a common standard

 

[Kamen Rider Blade]

Asus should make their proprietary eGPU port a common standard

Or we can get PCI-SIG to push the External OCuLink connector into mass production.

Spoiler: External OCuLink proposed design.

Then we can all benefit from having External PCIe connector that we can all use.

 

[edzieba]

Did anyone expect anything else??? it is PCIe 3.0 x 4 vs PCIe 4.0 x 16. That is an eight times transfer rate difference. Thunderbolt 5 with PCIe 4.0 might stand a little bit better chance with double the transfer rate of Thunderbolt 3.

For performance numbers that bad, it's very clearly not the bandwidth at fault: https://www.techpowerup.com/review/nvidia-geforce-rtx-4090-pci-express-scaling/28.html (TB3's PCIe 3.0 4x link is effectively the same as a PCIe 1.0 16x link).
From bandwidth limiting alone, only a reduction to 80% of stock PCIe 4.0 x16 performance would be expected. To see 50% or below indicates there is much more at play than merely bandwidth.

 

[Deleted member 14196]

🤣🤣🤣😂😂😂

junk. This is what you can expect from now on. With these ridiculous power requirements I don’t see things getting any better in the future.

this card is so stupid. what’s the next version gonna be like 2 kW of power required?!?

this has to be their worst design ever. The technology is moving backwards at this point.

people should vote with $$. Just say no. Save yourself $2000!!!

for that money, you can get a really nice designer dog. And the dog will love you, unlike this silly card.

 

[InvalidError]

The eGPU might perform better with an external monitor attached to it so output doesn't have to piggy back over TB.

 

[spongiemaster]

For performance numbers that bad, it's very clearly not the bandwidth at fault: https://www.techpowerup.com/review/nvidia-geforce-rtx-4090-pci-express-scaling/28.html (TB3's PCIe 3.0 4x link is effectively the same as a PCIe 1.0 16x link).
From bandwidth limiting alone, only a reduction to 80% of stock PCIe 4.0 x16 performance would be expected. To see 50% or below indicates there is much more at play than merely bandwidth.

For the HX laptop platform, there is no on die Thunderbolt controller. It's a discreet chip that has to compete with other data going through the chipset first. For such a bandwidth starved application like this, that would cripple performance. Jared did run a PCIe bandwidth test, and even in an ideal scenario synthetic test, both laptops had about 2.25GB/s of throughput, putting them closer to PCIe 3.0 2x than 4x.

 

[Spuwho]

All of these Thunderbolt docks have a x16 slot but are wired for x4. That works for some of the lower tier GPU's but not for any RTX models. These eGPU solutions are only designed to exceed a laptop iGPU in performance. Anything more is a mismatch of expectations.

 

[InvalidError]

All of these Thunderbolt docks have a x16 slot but are wired for x4. That works for some of the lower tier GPU's but not for any RTX models. These eGPU solutions are only designed to exceed a laptop iGPU in performance. Anything more is a mismatch of expectations.

It largely depends on how much bus traffic the games you play generate and what resolution you are pushing. You will obviously run into bus bandwidth issues if your goal is 1080p240 from the amount of scene data that needs to be rammed through the bus at high refresh rates but may still be fine at 8k60HDR10 where raw compute and local memory get a much stiffer workload.

 

[DaTruAndi]

What do y'all think how well it would work for pure CUDA use cases (e.g. AI inferencing or training)?