Another EU (Beta) tester in NL reporting 
Hey Filip. Potential Tester here, from Germany (Nürnberg). Würde gerne das Board Testen, bin als IT - Systemadmin “nahe am Fach”. Gutes Feedback garantiert. 
Viele Grüße,
Benjamin
Obviously you have enough people volunteering to be testers but I’m responding just to have something expressing interest and excitement about the project! Mid March, wow! I’d insert the futurama shut up and take my money meme if I had moderately less shame
1 Like
Hi Filip.
Great work! I would be happy to test this board. I ´ve got the first edition framework 16 and am looking for a powerful oculink egpu solution since the first days.
Greetings from Berlin, Germany
Martin
Hey everyone, thanks for wanting to be testers. I do specifically want someone with the AI 350 board.
@FWPL has the 7840HS, I have the 7940HS and @OVER_CL0CK has the AI 370 (and should hopefully get the board in the next few weeks since shipping from Germany to the US is kind of slow if you don’t want to rip a hole in your wallet). I ideally want a tester for each type of board. If no one with the AI 350 comes up in the next week or so I’ll most likely choose a random person just to have 4 people in total that tried the board before I proceed to a bigger batch. (But of course that means that it’ll be unconfirmed if the AI 350 board works)
I plan to open a survey of some sorts if the testers verify that the board works in their laptops where people will be able to enter their email so I can contact them and get a rough idea of how many boards I should order.
Anyway, I’m really looking forward to the testers hopefully confirming the board works and being able to provide a batch of the boards to the community. But like I said, I will unfortunately be temporarily moving countries in April, so I won’t really be able to produce any more boards until at least September due to the situation.
Unfortunately I don’t ever plan on selling these on a big scale. I will do a single batch if everything goes to plan in March and then if the demand is still really high, I will consider another in September if people are willing to wait that long.
1 Like
To explain a bit more about this issue. The current revision of the board has a small change in dimension that I have to walk back on…
Like it can be seen in this photo, the left part of the board is much higher than the area where I circled. I thought this would be fine when I looked at the dGPU board since it basically fills out that area, but unfortunately that does not seem to be the case.
When sliding the expansion bay back into the laptop, the laptop’s case gets stuck on the circled part. The tolerance is probably a fraction of a mm, but it is still an issue that needs to be fixed. It can be slid in if I push onto the board so that it gets pushed down a bit just before the shell is supposed to click in place. I did not have this issue with my expansion bay shell, but FWPL did.
I did this change since it gave me more space to have more ground around the PCIe lanes. Unfortunately I will be reverting back to the previous revision in that area. It shouldn’t cause any major issues to the signal integrity since the previous revision also worked without any hiccups.
4 Likes
Hey all, yes, I didn’t want to mention it so broadly, but yes, I can also confirm that this is an issue for me - it will need to be changed in the next revision. I ordered a new expansion shell I wanted to get in any case and also to do another fitment testing round with the board, but I expect it to be the same.
I just had an update from Aliexpress that the new cable went through customs and should arrive in the coming days. I already planned some time for extensive testing next week.
I’m still waiting for Germany to get their 
in a group and send my board from Filip. Been no updates in 2 days. My dock from Ali is at least in the US right now slowly making its way. Can’t wait to test.
Too bad, I got the Ai9 370 too
Hey Filip,
respect for managing to develop a functioning oculink adapter board, and thanks for sharing it with us. I’m sure it is a cool project to work on.
I’m interested the be a tester, if you are still searching someone. I’m from Münster, Germany and had odered my Framework 16 DIY with Ryzen AI 7 350 about 2-3 weeks ago, so I’m pretty new to the framework ecosystem. I bought it with the idea of adding an Oculink port in a mini project sooner or later.
I am currently using an Aoostar AG02 with a Radeon 6900XT via USB-C and am waiting for my Framework Dual M.2 adapter to be delivered so that I can pair it with an M2 to Oculink 4i cable. I want to test the differences and hopefully achieve a little more performance.
But I’m flexible to change my plans to add a 8i setup for even more performance 
I still have a spare ATX-PSU, so I just would need to buy a Oculink 8i to PCIe board and a matching cable (and maybe a 8i → 2x4i to play around, too).
~Alex
3 Likes
This might sound stupid, but I have no knowledge in the pcb thing. I soldered a few things, but I never designed or ordered a pcb. Could you give the noobs around here a tutorial for dummies on how to get a functioning board of yours, based on your github repo?
How did you do it, let it assemble by a company or just get the board and add some components ordered elswhere?
You could order the PCB with almost all components on it, but you would still need the screw standoffs (to be fair, these can also be 3D printed) for the 4 mounting screws and the screw nuts (these have to be soldered) for the interposer connector that you can’t get anywhere else but from Framework.
There’s also the fact that you have to buy multiple (usually at least 5 boards and you are required to pay for assembly of at least 2 if you go for JLCPCB). And there is also a base assembly price. The price per board gets lower the more boards you do, but the cost is not that pretty for a smaller run of the boards.
Hi all, I wanted to provide another update as the cable from Aliexpress has arrived and I am back at testing and trying to get it to work.
First of all, I received a second expansion bay shell and I was able to sand off a small part of the board at the location that @Filip indicated. It now fits much better, but still not fully. I cannot sand off any more, as it would compromise the fan connection on the board that are close to the edge. I went as far as possible and already double checked that the fans do still work. However, this means that, after sanding a part off, and after testing with two different expansion bay shells, the board does not fit well. Unfortunately it means I will have to order the next iteration too, but this is not a big problem of course. And I am happy to confirm that after sanding off about half a millimeter, that it fits already much better, almost as normal.
The second update is on the cable - and unfortunately it still doesn’t work. The error 43 fixer software sees the GPU after changing the script in the .bat file and running it, However, the notebook and device manager do not see it at all. The GPU fans do turn, so I think we are close.
What I will do now is to have another exchange with @Filip on possible solutions, and I believe it will require tinkering with the BIOS. I haven’t done this before and will ask him for a step-by-step guide on what to do, as I also want to get it right. Will provide an update asap, as I have time in the next days to fix this.
PS: error 43 fixer and Framework tool see Nvidia, so I think we are close…
3 Likes
Hey all, I am happy to confirm that it does finally work :), thanks to the help, once again, from @Filip . In the end, I had to reset the BIOS to optimal default settings (press F9, then F10 when in BIOS). Then, after resetting the BIOS, I had to re-run the scripted error 43 code. Will provide detailed testing results in the next days.
5 Likes
Here my first impressions from initial testing. Mind you, I have the following system specs:
Framework 16, AMD Ryzen 7 7840HS w/ Radeon 780M Graphics
Windows 11 Home, 25H2
64gb 5600 RAM from Adata
ZOTAC GAMING GeForce RTX 3090 Trinity OC (Deshrouded with Noctua fans)
Connected via OCuLink 8i
First of all, it is an absolute dream come true. I can’t stress that enough ;). And I have to thank @Filip once again for sending me the first board for third-party testing. I can confirm that it works perfectly, and that connection speeds are effectively much higher on OCuLink 8i, see below:
That being said, I also ran initial real-world gaming benchmarks and tested a first set of two games just now, and compared with OCuLink 4i (specifically the custom adapter from @Kyle_Tuck ).
The results are that the OCuLink 8i board does not deliver significantly better real-world FPS. It is the same, or even slightly worse than 4i:
Cyberpunk: 4i: 39-40fps; 8i: 38 fps
Far Cry 6: 4i: 67 fps; 8i: 64-65 fps
This is standing at exactly the same spots in game, looking at same landscape; and in the case of cyberpunk rerunning the benchmark several times in addition.
However, these results need to be taken with some context:
- firstly, I run games at 1440p (cyberpunk), and upscaled to 5k (far cry 6). Still, this remains a 1440p screen, and perhaps results would be better with a higher resolution screen.
- secondly, while I run my trusted 3090 and while it remains pretty powerful, it is an older GPU. So someone with a 4090 or 5090 may get better results, as there may be more throughput that can be used by the additional bandwidth that is now available.
Conclusion: I am happy to provide here the first feedback and first set of results, but it will be extremely interesting to hear the results from others with different systems and potentially newer GPUs. In my case, it may be the case that the GPU is simply maxed out and it doesn’t really make a difference.
A final thought: it may also be the case that the EGPU available is simply not as good as the DEG1, or that the signal integrity of the board of @Filip could actually benefit from some sort of redriver or signal amplifier.
But just to confirm the board works flawlessly, and it is a major achievement by @Filip . I felt honored to have received the first one, and I hope both my troubleshooting experience as well as my initial test results are useful. If I upgrade my GPU at some point, I will provide additional results. In the coming days, I will also run more games and get back again. And finally, looking forward to experiences from other testers now. Will be really valuable to have them.
UPDATE:
I could not really believe the results and tried again, different lighting situations/times in the games, did some restarts, a windows update, and again reinstalling newest drivers. And the results actually differ now:
Far Cry 6: 68 fps (so slightly better than 4i)
Cyberpunk: 41-44 fps (up to 10% better!)
So it may be that there was still a software issue after first setup. Shows you should not jump to conclusions right at the start. It seems to be fluctuating much more than 4i, but to perform better.
But do not expect “plug and play” ;). What I will do now is to take some time for testing over the next weeks and then report back with consistent average results.
7 Likes
When I was first building my own custom eGPU enclosure, I purchased a PCIe 4.0 x4 to M.2 adapter for low-stakes testing (if something went horribly wrong, I was only risking a cheap PCIe card and small SSD). That adapter (along with an M.2 to OCuLink ribbon adapter) was still just lying around, so I figured I could run some quick tests in my son’s computer over the holidays (late December).
These are the general specs. Ryzen 7 7700X, X870 motherboard, and RX 9070 XT.
I tested with the GPU installed in the desktop (running PCIe 5.0 x8) and connected via the PCIe to M.2 to OCuLink to AG02 (PCIe 4.0 x4). I ran a handful of benchmarks at 1440p. I didn’t have a lot of time (basically a few hours), so I didn’t figure out why the internal install was only x8 and not x16.
The 3D Mark PCIe bandwidth benchmark confirmed nearly 4x the bandwidth going from OCuLink to internal, but virtually every other benchmark ended up being margin of error territory (at best somewhere like 2% difference). I recorded the results somewhere, but now can’t find the sheet with the data. I decided not to make a post or report on that because the testing was limited to a handful of benchmarks (some 3D Mark tests, Cyberpunk 2077, and Shadow of the Tomb Raider, if I remember correctly). I knew people would likely be critical of my testing. To be honest, even I was skeptical of my results.
I agree that there are a lot of things that can be at play. Resolution, quality settings, CPU, RAM bandwidth/latency, etc all play a part, but since doing that testing my urge to switch to a x8 solution has diminished considerably. Don’t get me wrong; I’m still loving this x8 project (and may still try to acquire one myself), and I think the choices people have with the FW16 are amazing.
1 Like
My expectations were conservative in any case. I expected a moderate increase of around 5% additional fps at 1440p, as this would have been consistent with some benchmarks I found online and posted here in the forum some time ago. However, the relatively low performance you mentioned is also consistent with recent youtube videos testing pcie 5.0 (see here: https://www.youtube.com/watch?v=anoLAIEErHw), where pcie 4.0 and pcie 5.0 egpu performance is very similar at 1440p. The video does suggest, however, that it may be a different story at 4K, and especially with newest high-end GPUs. So it would be interesting to have someone with a 4K monitor and a 5090 to test the board. I may be the case that OCuLink 8i makes more of a difference in this scenario.
Otherwise it seems my results are consistent with yours. I do agree, however, that it remains a really interesting solution. And the board we have here is impressive (mind you, it works with a very cheap and simple EGPU and without any redriver or signal amplification).
So I agree with you, but looking forward to more testing still as well. If it really turns out that OCuLink 4i and 8i perform similarly, then I would be interested in the explanation why this is the case. I would normally expect that higher bandwidth especially with high-end modern GPUs should lead to more performance. Would be interesting to know why it is the same.
PS: we should also keep in mind that EGPUs matter. DEG1 and now soon DEG2 may just outperform the cheap simple stuff that is out there for OCuLink 8i in terms of EGPU.
1 Like
UPDATE:
I could not really believe the results and tried again, different lighting situations/times in the games, did some restarts, a windows update, and again reinstalling newest drivers. And the results actually differ now:
Far Cry 6: 68 fps (so slightly better than 4i)
Cyberpunk: 41-44 fps (up to 10% better!)
So it may be that there was still a software issue after first setup. Shows you should not jump to conclusions right at the start. It seems to be fluctuating much more than 4i, but to perform better.
But do not expect “plug and play” ;). What I will do now is to take some time for testing over the next weeks and then report back with consistent average results.
2 Likes
I will soon have my board from Filip and be able to do some testing as well. It’s almost here.
3 Likes
I’m really happy that someone else finally confirmed that the board works as well. So I’m looking forward to more benchmarks from others! If everything goes alright in the next ~2 weeks, I’ll be ordering more OCulink connectors to JLCPCB to do a bigger batch that should definitely have no issues, since the only issue is the small change in dimension!
4 Likes