Well it depends hardly on your Games and scenario. If one core is pegged to 100% your cpu is your bottleneck, also when one ore more cores are sitting above 90%. If your GPU is above 90 or pegged to 100% its your gpu. If its neither and you are not getting “enough” Frames it can be the engine of the game/ memory optimization OR the bandwith to your GPU thats the hardest Part for an eGPU Setup. Its not really obvious. The worst is when using a wireless VR Setup or the internal Display with an eGPU because the whole datastream has to be backfed to the device, but there is no obvious marker on the telemetry. For my example using flight simulator with vr with my old setup with usb4 and rx6800xt the gpu and cpu “idled” below 50% usage and i had barely usable frames. With my widescreen monitor the GPU and CPU had been utilized above 90% and it was flue t around 60fps it also was better after i upgraded to the oculink setup to use with vr.
Coincidentally after asking my question earlier about the local nvme slots on main board; if you haven’t seen this yet check it out..
“Secondary Storage options (M.2 2230 socket)
Choose up to 2TB WD_BLACK™ SN770M NVMe™ M.2 2230 SSDs or bring your own
Note: The secondary storage interface supports x2 PCIe 4.0. SSDs with x4 PCIe 4.0 will run at slower speed.”
This helps explain the lane configuration. The HX 300 series lacks 4 lanes found in the 7x40HS. Dropping the 2230 down to two lanes is part of it. However, the HX 300 series pick up an extra USB 3.2 gen 2 port that the 7x40HS may have had to sacrifice lanes for.
But the dual M.2 expansion module doesn’t lose anything, so that’s good.
Doesn’t the FW16 have 2 x USB 4 & 4 x USB 3.2 ports already? Apart from knowing that the HX 300 looses 2 lanes on the 2nd M.2, has there been anywhere else stating where the other 2 are lost/affected?
The 2230 m.2 slot was reduced to 2 lanes instead of 4.
There is a little more descussion on PCIe lanes from #20 on:
Transcript of the Q&A stream with CEO Nirav Patel
Alright, thanks! I mainly asked as I remember that also the amount of cores matters (I believe 4 core CPUs are inherently a disaster nowadays if I am not mistaken) and I used to play Cyberpunk (admittedly a very CPU heavy game) on a 6 core Processor, and once I upgraded to 8 core it went much more smoothly. But yes, I understand, so “better” CPU does not automatically and not always per se help, even a couple of fps.
So, I finally got everything working! I hit a few dead ends, so documenting my experience here for others.
Like many, I was waiting for Josh Cook’s 8i bay. So, the first batch of items I bought was:
- a case (I wanted the whole setup to be enclosed)
- that one 8i dock everybody seems to get
- an 8i cable (1m long; this was needed as you need quite a bit of cable length just to get the cable out of the case)
- The graphics module interposer
Basically, this setup:
When this thread popped up, I figured I might as well buy a GPU since I could use the nvme expansion bay if Josh Cook’s board did not materialize. I went with an RX9070XT (I’m on linux and don’t want to deal with nvidia).
Since some people in this thread mentioned that they got this chip working with a 8i → 2x4i cable, and since Josh’s board did not materialize, I figured I might as well give it a go.
The second batch of items I got:
- The NVME expansion bay module
- An M2 to oculink ribbon cable
- An 8i → 2x4i cable.
I also 3D printed the IO cover (I needed the one with 4 holes) and the plugs.
Unfortunately, this set up did not work. The GPU got power through the board, but nothing happened when I booted my laptop. I could also not see the GPU using lspci.
I was tempted to get the DEG-1 people in this thread keep recommending, but I don’t like the gpu being out in the open and I liked the aesthetics of the case I got. Thus, I ordered this dock (linked earlier in this thread and just happened to work with my case) instead. It came with a 50cm cable, which sort of works since the dock’s oculink port is in a better location. I would have preferred a longer cable for my desk, but signal integrity might be the reason that my previous set up did not work, so I didn’t want to risk it.
Some messing around with screws later, a little hail Mary and … it works! The GPU was detected and can now be used with no issues. I ran the UNIGINE Superposition benchmark and got >150fps average at 4K. Performance in cyberpunk was less good, but that might have been me being a bit too ambitious with the settings.
While the performance seems to be nice, I would prefer to have a complete solution that allows me to plug the integrated display of the laptop into my eGPU. But for now, I have a workable solution; after months of waiting, I only need to figure out which games I want to play :).
Summary of the complete working setup:
- On the laptop side:
- An oculink 4i cable (50cm)
- On the GPU side
- Oculink → PCIe Dock
- Case
- RX9070XT
- An old power supply from my desktop
What ended up not working:
- The 8i dock
- 8i→2x4i cable (1m)
I’m not sure if the dock, the cable, or some combination of both is to blame here.
Trying to understand here.
You can break 8i to 2x 4i, but not backwards.
What did you get working? 4i without DEG1?
Some people in this thread reported that it worked: link. Since I already had the 8i board, I figured I could try it out. It didn’t in my case.
Yes, plain 4i with the dock listed in my post.
I have been putting of the 2x m2 board, you are using that?
Yes, this thread is about having oculink working with the m.2 expansion bay module; I guess you might have ended up here from me quoting you without being aware 
.
So my setup is nvme expansion bay module → m2 to oculink ribbon cable → oculink 4i cable → oculink to pci dock → gpu.
I’ve been searching all over and can’t seem to find a good solution for this. I am curious what 3D printing lab you used? Was this something like a website or just some printing service near you?
Just bumping the thread to check on update/status to how this was going for you?
Well, it’s funny you ask. I have been debating whether to post anything or not. Even though I don’t really believe in signs, I’ll try to take your specific query as a sign that maybe I should post an update.
After considering various ways forward, and understanding that I would be taking on (what is to me) significant financial risk, I decided to order the initial 10 pieces. I was concerned about the scenario in which they didn’t work and having random people “on the Internet” blame me for wasting their money, but I still really wanted this to move forward. The manufacturer makes a lot of different style adapters, so that figured into my decision as well.
So…
I have not been successful with getting them to work. I have tested 3 of the 10 with no success. I have a custom made dock built using a cablecc host adapter and an Aoostar AG02. I installed an SSD in the slot to make sure the Framework board wasn’t the issue (the SSD was detected fine). I wasn’t really surprised by the cablecc adapter not working, but I was more hopeful about the AG02.
I contacted the manufacturer about their testing (which they told me they would do) and they replied:
We tested them with GPD G1 graphics card and motherboard m-key and it work fine.
Obviously, I’m disheartened. It is possible they might work with the DEG-1, but that would mean sinking even more money into this for yet another gamble. It would also mean these would be limited to a fairly specific hardware combination (though the DEG-1 seems to be the most popular choice in this thread).
Don’t suppose anyone would be willing to help finance a DEG-1? 
What a journey it seems to have been for you. But I certainly do appreciate the update. (A sign he says 
) Sorry it doesn’t seem to work for you though. I don’t understand how they said they tested and it works, but it’s not working for you. I have the DEG1 that I am currently using for my set up. I’d be willing to test it 
Ugh, was also hoping for some success here too as I have an Aoostar AG02 [800W] and looking to get off the TB4 connection for it. Looking through all the solutions, this one seemed to be the most viable instead of dealing with ribbon cables and risking damaging them while bending. I think I saw the recommendation to use the non-ribbon cable types to avoid that from happening, but unsure what item out there exists that is proven to work nicely with the given strategy of a custom 3d printed back panel.
Would be great if Framework simply came up with their own solution that would work here. I don’t see oculink going away in the future as there is simply no better way to connect pci-e via cable otherwise.
I am not an EE, but just in case this thought is helpful - for differential pairs in general and PCIe in particular, I’ve heard that you have to make the length of all the traces in all differential pairs the same. From your photo I get a vague idea that maybe the traces with the outer bend are maybe longer than the traces on the inner bend. Couldn’t say for sure without a full schematic, and I am not an expert, but maybe it’s something that’ll help you.
Edit: another thing is you haven’t mentioned what GPU you’re testing with. If it’s PCIe 4 card, try testing with PCIe 3. Maybe not even a GPU, but and other PCIe 3 card you have. Third version is more tolerant for signal integrity (I couldn’t get a PCIe 4 GPU to work on my setup) and knowing that any PCIe at all works on your card might help narrow the issue you’re experiencing.
Edit 2: Another wild idea that just came to mind is maybe PCIe is tolerant to having fewer lanes connected? Wonder if you could downgrade the connection to x1 by cutting all pairs except first in each direction. If this is possible, this way you eliminate the trace length requirement (maybe) and if it works then, you again have something to go by…
@Kyle_Tuck here we go. Gmanny said the problem. I attached a link to an pdf about pcie routing. https://www.ti.com/lit/pdf/snla426
Another good point is mentio ed here for guidelines: https://www.andwinpcb.com/high-speed-pcb-routing-rules-for-pci-express/
But most just discuss interpair length matching. I would go the route and do all lanes length matched. Its all about timing, impedance and crosstalk.
With those m.2 to oculink wired breakouts, every wire is the same length and they are shielded. Pcie 4 is especially sassy about timing, different wire lengths are hitting on the timings, as every of the 4 channels have to be within a same range.
Although it isn’t a laughing matter… that is why you see The Wiggles on boards when attempting to length match around bends.