OcuLink Expansion Bay Module

[OVER_CL0CK]

Ok everyone, I got Filip’s board in and was able to get it to work on my AI 370. It definitely works and shows the difference between 4i and 8i both in GPU-Z Bus Interface and AIDA64 data speeds respectively. I will try and keep this as short as possible as images make it quite large.

GPU-Z 4i:

GPU-Z 8i:

**To FWPL’s point, this seems to be a very game specific variance as I tested a few games. Some did slightly better, some did worse.**

For starters, everything was done on the exact same version of Windows 11 Pro 25H2 with the AI 370 Mainboard, 32GB of RAM, Power Mode: Best Performance, using AMD 9070 XT with Adrenalin 26.1.1 @ 2560 x 1440 on a 27inch external Samsung monitor with a 165hz refresh rate. All games were at ultra settings and had all frame gen turned off (Adrenalin had it turned off too). The only thing active (if available for the game) was FSR 4 upscale.

---

~Red Dead Redemption 2: (Better)

This game is the only one I had on hand with a built in benchmark at present.

4i Benchmark:

Minimum: 15.6114 FPS

Maximum: 158.798 FPS

Average: 121.089 FPS

4i In Game:

Lowest: 100 FPS

Average: 112 FPS

Peak: 118 FPS

8i Benchmark:

Minimum: 15.193 FPS

Maximum: 164.872 FPS

Average: 125.086 FPS

8i In Game:

Lowest: 102 FPS

Average: 114 FPS

Peak: 120 FPS

---

~The First Descendant: (Better)

4i In Game:

Lowest: 63 FPS

Average: 77 FPS

Peak: 81 FPS

8i In Game:

Lowest: 65 FPS

Average: 84 FPS

Peak: 91 FPS

---

~Escape From Tarkov: (Worse)

4i In Game:

Lowest: 109 FPS

Average: 114 FPS

Peak: 144 FPS

8i In Game:

Lowest: 76 FPS

Average: 98 FPS

Peak: 110 FPS

---

~Starfield: (Better)

4i In Game:

Lowest: 74 FPS

Average: 77 FPS

Peak: 81 FPS

8i In Game:

Lowest: 74 FPS

Average: 80 FPS

Peak: 86 FPS

---

~3DMark: (Worse/Better Respectively)

4i Steel Nomad:

Score: 6986 / 69.86 FPS

8i Steel Nomad:

Score: 6918 / 69.18 FPS

4i Time Spy:

Score: 26829

Test 1: 177.96

Test 2: 151.49

8i Time Spy:

Score: 26965

Test 1: 180.32

Test 2: 151.22

---

I know it was a short list, and these were all down and dirty as soon as I got the board working and run in the same order from 4i to the 8i swap. I am not a tech expert, so I don’t have the answers as to why this stuff is showing some worse data. What I can definitely say though is my GPU was able to actually reach 99-100% utilization on the 8i board whereas it was never able to do that on the 4i. Averaged high 80% with 4i. You’d think it would mean more of a difference, but not as drastically as it would seem. More GPU utilization, the less bottle-necking occurring, which is definitely a plus side on the 8i factor as twice the lanes are being utilized therefore more headroom.

---

Final Thoughts:

-Does it work? Yes, it certainly does. On now confirmed both the Framework 7X40 series and 300 series.

-Is it WORTH it? Your mileage may vary. If you do not have the 4i variant, 8i may be for you as it cost about the same for the 8i, just not as nice a set up as the DEG1 (AliExpress version is bare bones board). Me personally? No, not compared to getting the DEG1 for the 4i with KyleTuck’s adapter and then spending additional for the AliExpress Board and cable. As others have said on other forums and numbers prove in this thread, the gains are only roughly 3-5% whereas it seems some are loss (this could be a potential handshake issue somewhere).

---

I would like to say in closing, I did this to see for myself the potential of 8i (and to help a fellow developer), and while the gains do not seem to be anything drastic, I do want to applaud @Filip for creating (to my knowledge) the first successful 8i board for the Framework Laptop. It was a great privilege to be a part of the testing community and get to share with you all. Maybe as time progresses, I can find additional tweaks to see if there are differences in performance, but for now this is my review.

[piLovr]

That’s quite interesting, I really thought the difference would be more noticable. Yes, the 8i to 16i diff is probably low, but I thought 4i → 8i would be a bigger jump.

The following is jot relevant to you, but might be for anyone:

I have (link down the post) an aliexpress board, of which I thought it would be 4i, Gpu-Z shows 2i, which is weird (or I just ordered bs).

I have a pretty similar setup to yours, 32’’ 1440p 165hz monitor, I’m gonna add some benchmarks (probably tomorrow) of RDR2 as soon as I’m home, maybe someone cares about it xD

I’m very curious how big the 2i → 4i jump is, so maybe I’ll switch to the DEG1 actually.

edit: I got the 9070xt too

https://a.aliexpress.com/_EyrI7j6

[James3]

@OVER_CL0CK
That is a nice set of tests. I think it demonstrates how little games tax the pcie bus. They offload so much to the GPU that the pcie bus is not used much.
Latency does make a difference in games. So 4i is better than the higher latency usb4/thunderbolt. There is very little latency difference between 4i and 8i.
Applications that would use the bus more:

1. llm model load times.
2. using the gpu for compute, where large quantities of data are passed to/from the GPU such as large matrix operations.
3. pcie network/infiniband cards showing data throughput at 100G.
4. clustering compute, using infiniband between mainboards.
5. using docks that have more than one pcie slot. With a mix of SSD and gpu or tpu.

[Filip]

Great to know that almost all of the boards have been verified to work. There’s just the AI 350 left, but that should also be tested in hopefully a few days.

I did just order another batch of what should be the actual final revision of the boards. I made the trace corners more rounded instead of using hard 45 degree bends which, even though the signal integrity seems to be fine, should help further improve it.

@Axel_Boersma

This Forum might work, but difficult to keep track of everybody. And we need to organize getting addresses, payment etc…

Yes, I’ll see how to deal with this in the coming days. Most likely a Forms page in order to see how many boards should be ordered.

Just keeping my fingers crossed that I am able to get, test and ship all of the boards before my move date at the end of next month.

EDIT: It might be a good idea for people to organize groups for cheaper shipping. Especially to the US due to regular post service only allowing gift shipment due to the various changes happening there for incoming international shipments. I will have to ship the boards via the next cheapest option (UPS) and the shipping cost comes to around $45 for that. Other places should be cheaper since I can use regular post service (should be around $30 for places like Australia, half of that for EU), but if someone wants faster shipping it’ll most likely be around that amount as well.

[Filip]

I have just written a simple Forms document for applying to a waitlist. I’ll probably close it in about 2 weeks and then see how many boards will have to be ordered in the end.

Framework 16 OCuLink 8i Expansion Bay Board

[Filip]

@Filip

Just keeping my fingers crossed that I am able to get, test and ship all of the boards before my move date at the end of next month.

For anyone interested, I’m still just a student, so I’m doing a whole semester abroad in Tokyo for about 5 months starting from April. I might bring a few boards with me there in case someone wants to personally pick them up if they are in the area. I doubt there’ll be anyone, but might as well make the offer in case the shipping costs are too insane for some people .

[Filip]

It’s puzzling as to why Escape from Tarkov saw such a bad performance drop. I do wonder if its actually related to signal integrity or if it some odd driver issue when swapping the cards.

I wonder what would happen if you were to switch to x4 configuration on my board.

[OVER_CL0CK]

Hard to say. Due to different spawn points on maps, I made sure to play a “practice raid” so I could get the the exact same spot in the map and get read outs. I was quite puzzled by this one as well. I know that game is quite CPU heavy and didn’t expect a decrease in the gpu performance, but maybe there is something I am missing. Quite honestly, I thought about dropping that game from the test results because it really isn’ the most optimized, but I wanted something demanding in my tests.

[Michael_Butash]

I suspect from game to game you’re going to find wide variances based on the engine, level of optimization, even just the type of games for what you can expect. The issue comes down to what data is being processed on the cpu vs gpu, and shuffled between them constantly. I’d be curious to monitor the pci-e bus during transfers along with cpu, gpu in linear graphs side-by-side to see which is bottle-necking first for each.

You should be able to see the rx/tx on the bus with “nvidia-smi dmon -s t” or “amd-smi monitor --pcie” for either windows or linux or with Adrenalin overlay for amd (so chatgpt says, not a windoze user here), probably do some interesting graphing of the data in graphana/prometheus if you wanted…

As @James3 said above, ideally throw in a 40/100/200 gigabit ethernet/infiniband nic in there with a packet generator to really see what sort of linear performance you’d get out of the bus, but need two of them to run back to back. I have a TB3/4 to 10gbe nic SonnetTech adapter that can do the expected 9.3Gbps or so with iperf, did always want to pop in a 40gbe nic to see TB4 cry some…

[FWPL]

That would be absolutely amazing. My own plan was to go to a 3D printing lab with the files/dimensions of the FormD T1 as well as the DEG1 case and literally ask them to develop a 3D printing file. But if you could have a try first that would be amazing and no problem to wait until March of course!

[FWPL]

Great to see it worked out! Just a general point, in my case some of the games also performed worse at first, and then literally a day or two later simply performed differently, and better. It was very strange. To me, it almost seemed as if the EGPU and my system had to do a sort of “handshake” over time. I know it sounds strange, but I do confirm several games first had worse FPS, and this changed after some time. My final review was a couple of days later after constant testing. Perhaps you can recheck the game with worse FPS, just to check if it is maybe the same with you.

Finally, I also wanted to share the real-world fps differences of the two games I tested the most:

Cyberpunk: 4i: 39.5 fps; 8i: 41.5 fps

Far Cry 6: 4i: 62 fps; 8i: 67 fps

[FWPL]

@Filip

It’s puzzling as to why Escape from Tarkov saw such a bad performance drop. I do wonder if its actually related to signal integrity or if it some odd driver issue when swapping the cards.

I think it might be signal integrity too. After all, our EGPU we are using is literally the most basic I have ever seen to be honest. I wonder if 8i would be much better with a serious EGPU with signal amplification.

And another question: do we have a tester with a more powerful GPU (4090, 5080, 5090)? It would be interesting to see if that would make a difference.

[Filip]

@Andrew_Shaffer

I did pick up the 8i-eGPU board referenced here for experimentation; haven’t gotten the cable yet, need to get that ordered. I’d be happy to take a whack at remixing that case for this board and putting it up on Printables.

I would absolutely love to see a nice enclosure or at least a mount for the board. My current solution is alright, but it was quickly put together to hold the board and an SFX PSU in place.

@FWPL

I think it might be signal integrity too

We’ll see with time. I might try adding a redriver to this setup in the future. Might be another nice modification to the project. Ideally we would want a retimer, but the complexity of those things scares me . But I still do not see signal integrity being worse than the M.2 board solution.
I would love it if someone with an actual oscilloscope could check the eye diagrams of this and compare it to the M.2 solution. But 16 GHz oscilloscopes are insane in pricing so there may not be anyone with the FW16 and all the boards and cables on hand.

[Andrew_Shaffer]

Thinking I’ll start by putting together a dimensionally-precise (as best as I can do with my calipers, anyway) model of the eGPU board we’re using. Should be straightforward to take that and mod adapters like one for the case you want to attach it to, plus lets others derive their own mods for mounting the board more easily. I’ll be sure to post here when I have that first model rigged up, then check in to get dimensions / mounting points you’re working with in your FormD case.

[Andrew_Shaffer]

I always expected I’d design something for my board + PSU + GPU anyway. Only seems sensible to reshare that back with the community here.

I’ve got my cable on the way, and I’ll be able to test two extremes eventually when your 8i board is in production - a 1050 Ti and a 7900 XTX.

I should note I run Nobara Linux on my 7840HS, so it’s going to be… interesting, shall we say, to see if/when I get it all working. I don’t mind slotting in a second SSD to test with Windows too though, would just be lower on my priority list.

[OVER_CL0CK]

I find it quite interesting and wonder if AliExpress changed their board model. Because I got a 3d print from his board design (Filip) and it did not match mine (mine seems longer), so I frankensteined this together as I have no official mount/platform. Might want to compare what dimensions people have?

[Andrew_Shaffer]

Very interesting… Not terribly surprising, either. Might be we want to check for model/revision models on our eGPU boards and compare overall dimensions to see what variations are out there - especially if there are varying mount points.

I’ll pull my board out here in a bit and see what it says/measure it out and share here. Starting a running list of variants seems like a good idea.

[Lou_Cipher]

This is fantastic! 4 hours ago, I’d never heard of Oculink, discovered it because amazon randomly recommended an nvme to oculink adapter to me, and I got curious. 15 minutes later I was reading this thread. Now I’m on the waitlist. Got both the gen2 370 fw16 and the original gen1 7940hs here.

I noticed everyone seems to be testing this on Windows. Anyone try this out on Linux yet, or will I (hopefully, if I the waitlist goes my way!) be the first? I’m used to being the linux lab rat, but I’d be curious if it’s been tried yet.

[PSierra117]

Works just fine on linux. Had used oculink on my ssd bay long before this adapter on linux.

[Overheat]

Hi everyone,

A quick update:
The 8i board and cable finally arrived from Aliexpress yesterday, so I was finally able to test Filip’s board.

Unfortunately, I haven’t yet managed to get my Radeon 6900XT and an old Nvidia GPU to run flawlessly on my FW16 (AI 7 350) in x8.

Neither a BIOS reset, battery disconnect, rollback from BIOS 3.04 to 3.03, reflashing the firmware on the Oculink board, disabling PCIe Dynamic Link Power, nor replacing the ATX power supply helped.

Unlike the Nvidia graphics card, the AMD is recognized in Windows, but still only runs in x4 mode, causing the system to lag and the CPU to get stuck at around 0.7Ghz and no longer boost. Interestingly, as soon as I disable the GPU in Windows Device Manager and restart, the problems go away, but then of course the GPU is unusable.
However, when I switch Filip’s board to 1x4i (flashing the other firmware is sufficient) and restart, the problems disappear and I can use the eGPU without any issues.

After many attempts, GPU-Z now reports the use of 4.0x8, but the remaining symptoms and really poor performance remain.

Under Linux (Fedora), both GPUs are listed in lspci in 8i mode, but there is an error (Unknown header type 7f) under lspci -vv with the Nvidia, and the performance is also poor with the AMD GPU.

I’m currently at a loss as to what else I could test to get rid of the problems. Since it already worked with OVER_CL0CK Framework with AI 370, I didn’t expect any big problems.