USB4 eGPUs in 2025, is it worth it? What is the limit?

[Agall]

TLDR: eGPUs via USB4 are not bandwidth limited at 40Gb/sec in any significant way, the limitation has been and still continues to be the limitation of the eGPU host computer's CPU performance. Same as usual, IPC/frequency/cache, but also potentially its ability to maintain stable boost clocks via firmware/VRM. Is it close enough to 'zero compromises' for even a seasoned hardware enthusiast? Yes.

 

Some background:

I first started testing eGPUs back in 2016 with a Razer Core and Razer Blade Stealth. The Stealth wasn't enough, so I bought the regular Blade instead, which I daily drove for a year with a GTX 1080. There were severe input latency limitations that I was able to measure at ~50ms.

 

Second time I tested this was in 2020 with a Lenovo Yogabook, which I daily drove for ~4 months with the same Razer Core and GTX 1080. Input latency wasn't severe enough to be a problem, but I still found myself finding the setup insufficient.

 

As of yesterday, I concluded a ~6 month period of only using an Ally X and USB4 eGPU connected RTX 4090. I used a 30W manual profile with fans locked at 85%. Whenever I'd 'forget' to set the 85% fan lock, the performance different was instantly noticeable in gaming.

 

Testing

 

Between these three experiences, the most recent Ally X configuration is by far the closest I've gotten to 'no compromises' by a mile, so close that I would've had no problems sticking to it. So why did I eventually switch back? Spoilers, I get bored and like testing things about as much as I like gaming itself.

 

The only major compromise I'd run into is in certain scenarios in Warframe, the framerate would drop to 60 fps. This lead me to theorycraft why, so, I did what any reasonable person would do and spent a ton of money buying hardware to test it.

 

USB4 + Ally X, so what's the limitation? As it was in 2016 and 2020, a CPU limit. How do I know this? I bought a 8700G and 9800x3D just to find out.

 

I won't have crazy graphs and tons of quantitative data to demonstrate this, since I operate in the realm of practicality and the experience itself. The nature of the limitation being around a live service game creates limitations in variable control, but as someone who's played Warframe for over 11 years and has tested dozens of build configurations using the game, I've been able to eliminate most of them.

 

I had three significant scenarios and limitations to test. Baseline hardware includes the same motherboard, stock cooler (yes I put a stock cooler on a 9800x3D, please don't Su me), 6400MT/sec kit of RAM, and my RTX 4090 at default settings for all of these tests.

 

Cassini, solo SP, staring at wall in small spawn room tile, 1080p low DLSS quality

-Ally X 30W and fans at 85%, ~400 fps

-8700G 35W 4.1GHz (to match Ally X as best I can) USB4 ~500 fps

-8700G 65W USB4 ~600 fps

-8700G 65W PCIe 4.0x8 through 3.0x4 ~650 fps

-9800x3D 3.8GHz USB4 ~850 fps

-9800x3D ~5.2GHz (no limits) USB4 ~1200 fps. Arbitrary framerate limit on USB4 eGPUs? I say not...

What I conclude from this is that the Ally X's 278mm^2 Phoenix APU isn't 1:1 with the 8700G's same die, even when locking the clock speeds and TDP to be close. The theory in this is that the Z1 Extreme is binned for the iGPU, the 8700G is binned for the CPU, as well as the VRM/firmware not being designed for high sustained boost clocks. Obviously by these values, USB4 isn't a significant limitation here. I'll add the 9800x3D with PCIe 16x later, that's the config my system is in currently.

 

Second scenario where I was explicitly looking at minimum framerates is in the new 1999 zone, Scaldra survival mission, solo SP. In this scenario, the Ally X would drop to as low as 60 fps for longer than a second or two at my normal 4k medium quality DLSS. When I tested this with the 8700G at 35W, 4.1GHz, and via USB4, I was able to recreate the frame drop but to about 65 fps. Overall, it was more stable on the 8700G, but still a limitation. What was evident is that its a hard CPU wattage and L3 cache limitation after testing the 8700G at full wattage and via PCIe, since I didn't experience this to such a severity with the 9800x3D in its most limited setup (which would only drop to about 100 fps).

 

Third scenario, and by far the most limiting with the most interesting results that surprised me. 4 player (host or client at <100 ms) Omnia SP Void Cascade.

-Ally X, regularly dropping as low as 60 fps, would maintain ~75 fps.

-8700G 35W 4.1GHz USB4, ~140 fps minimum

-8700G 35W 4.1GHz PCIe 4.0x8, ~140 fps minimum, no difference from USB4

-8700G 65W PCIe 4.0x8 ~160 fps minimum

-My previous 7950x3D rig would maintain ~240 fps in this scenario for reference, which is where I came from before the Ally X

-The same rig with CCD1 (aka non 3D v-cache) would maintain ~160 fps.

-9800x3D 'no limits' PCIe 4.0x16 ~300 fps and likely to only get better since its running a stock cooler

What I conclude from this is the same as the previous test, a hard CPU limit, but I was surprised how well the 8700G did. The theory being that 16MB vs 32MB isn't that different in scenarios where L3 cache is the limit, but 96MB is plenty. Regarding Z1 Extreme vs 8700G, this is where my theory behind boost clock stability and VRM comes in. Even with the CPUs at the same clock speed, I don't see why this wasn't comparable outside of variables like firmware/VRM differences or binning. It would make sense to me that the Z1 Extreme is binned more for its iGPU and 8700G more for its CPU given their expected use-cases, but that's far larger than I would expect that difference to make.

 

Outside of someone who specifically wants to run this kind of setup, there's practicality here, especially with the upcoming Z2 Extreme refresh. I would argue that someone already with a capable GPU and power supply, but might've skipped a few CPU upgrades and is looking at handhelds, might be better off buying a USB4 dock for $150-$200 and build an eGPU dock using that existing GPU+PSU. The advantages being portability and simplicity, the major disadvantages being upgradeability and value, since baseline this setup would get close to $1000 for just the handheld and dock itself.