Im sure there are many users which would like to replace their desktop and don’t wan’t to use the offered dGPU. Therefore we need oculink for eGPUs.
So instead of having a bigger expansion bay with an “internal” dGPU, we could have similair sized expansion bay, which offers an additional battery AND an 8i oculink. Because it has already 8 PCI-E (4.0?) lanes, this is basically perfect.
Im sure a lot of people also would like to have a battery expansion bay so the idea is: One expansion bay with both inlcuded. So that means the potential userbase is bigger on both sides, which makes the bay more worthable.
Alternatively there could be an expansion bay with only an oculink 8i (or also offers 2 different expansion bays: one with only oculink 8i and one with battery + Oculink 8i. For people, who dont wan’t to sacrifice additional space. But because one Expansion bay is enough for the start, my idea is to combine it with an anyway wished battery-expansion bay. So it would be a win-win situation.
Is there any chance, this could happen?
Maybe you ask now, why you need an eGPU ? The 16" model does have a dGPU and will have GPU upgrades in future. Well, there are several points. The first reason might be, you wan’t to have “normal” graphics cards, which means you have free choice of a graphics card. For example you can buy a RTX 4090 or RX 7900XTX or what ever. You are not bound on 1-2 options which maybe have not enough perfomance, you want. Or maybe you are bound to nvidia because of software or what ever reason. You just have completly free choice. And second reason is, it will (to be honest) be cheaper, because its open market and it is just more expensiv to design such an attachable GPU. Third reason is, there will probably never be a RTX 4090 with several hundrets of watts, because a laptop just can’t deliver that much of power. So yeah, if you want such a card, we really need oculink 8i.
I think that is an awesome idea. An expansion module with cooling+more battery+eGPU connector and a proper desktop GPU enclosure. Majority of laptop gaming is done at the desk where you also have your gaming mouse+keyboard, large monitors, headset, etc so there is no reason to take your dGPU with you in many circumstances. I would consider going for that if it was available.
The issue is, is that people will want to have so many different combinations it’s hard to decide on a specific model/design, also the battery will not fit in a standard expansion bay shell unless you go downwards (Z axis) due to the fans.
Fair question. First one is, theoretically, the dock can be more cheaply engineered and produced - Occulink’s a pretty specialized standard.
The other one may hinge on how much traffic does a big GPU like the AMD 7900 xtx, nVidia 4080ti or 4090 actually put through it’s PCIe channel… I’ll have a little poke around. Anyone else who knows go ahead and chime in.
People choose oculink as it was a connector that has been used to have more bandwidth, as gpus typically need much more than thunderbolt can supply (TB4 supplies 40gb/s whereas you can get greater than that via Pcie through oculink.)
So OP has a point, that review of a 4090 - consumer GPUs don’t get any heftier at time of writing. It saw an average 2% performance dropoff over PCIE x16 gen 3, which is half what the Gen 4 test can throughput and somewhat slower than the TB4 cable is rated for.
Ergo, with the (probably) highest throughput GPU consumers would ever be crazy enough to buy, going faster than TB4 bought… a 2% gain.
This doesn’t match up with results I’ve heard elsewhere about eGPU performance. How interesting. Anyone else want to chip in? I better get back to work.
*Edit: I was wrong. Gbit vs Gbyte, makes a big difference! See others below.
According to website you have shared, the average loss in performance is 20%.
Thunderbolt offers a PCIe 3.0 x4 interface, which offers a bit less bandwidth than a PCIe 1.1 x16 slot (3.94 GB/s vs 4GB/s). Whereas increasing that to a 3.0 x8 slot will only cost you about 6-8% in performance.
There are also some productivity focused applications, which rely on large amount of data to be supplied to the GPU. Doubling the speed of the transmission would nearly double the speed of those applications. But I doubt anyone would run them via thunderbolt/USB4 anyway.
That’s just NOT true. Ok, i see there is a lot of misunderstandig what the difference of thunderbolt and oculink is:
First of all: It is not only about bandwith. eGPU solutions with thunderbolt often just doens’t work properly, like giving errors, crashes, random diconnetcs, half lane speed and so on, while with oculink, it just works! So bandwith is not all! Compatibility is bad with thunderbolt! So actually i could stop talking further, because doesn’t matter how good thunderbolt’s bandwith is, it just often don’t work porperly and oculinks does. Maybe this is because oculink is basically PCI-E while thunderbolt is thunderbolt, but i don’t know and actually it is not important because oculink just works and thunderbolt not.
(Probably also one of the most important points): With oculink you could build your own setup which is quite cheap. If you use thunderbolt you need one of those proprietary and very expensiv docking-stations…horrible and no one who loves frameworks should like that actually.
Second (bandwith): Thunderbolt 4 offers only 32GBit/s which is 4GByte/s which is same speed as PCI-E 4.0 X2 or PCI-E 3.0 x 4 or PCI-E 2.0 x8 or PCI-E 1.0 X16. So this is like horrible. And also important is, that thunderbolt does not support PCI-E Gen. 4 at all.
Oculink 8i offers same speed as 8 PCI-E 4.0 lanes, which is like a fully PCI-E 3.0 x16 ( “normal” oculink ports have “only” 4 lanes, so not 8i not 8 lanes). So oculink 8i offers 4 times bigger bandwith. Btw: thunderbolt 5 (not out in any device yet) has 80GBit/s which is 10GByte/s. That would still be less than the Frameworks 8 PCI-E lanes offers or oculink 8i. But maybe only 64GBit/s is available as data because thunderbolt 4 offer 40GBit/s and 32GBit/s data. But i dont know.
Third: The losses of today will be even bigger tomorrow. So you wan’t to have biggest possible bandwith, and that will be oculink 8i for loooong time (because oculink will have even compared to thunderbolt 5 60% more bandwith). So and Framework is about longevity and upgradeable (and therefore sustainability). So therefore you need the biggest bandwith, you can get to make sure, your device will not be e-waste in 5 Years. This laptop is meant to hold longer than 5 years (at least for me). I mean, even oculink 8i has “only” the same bandwith with outdated PCI-E 3.0 x16.
Fourth: The benchmarks from techpowerup showed above are nice to have, but (sadly) like techpowerup often does, they only show average fps, which does not show the real experience, you will get.
Fifth: 16" Framework laptop is not only for gamers, also for people who do heavy workloads and then, you even need more the biggest possible bandwith. So don’t forget that.
Sixth: Resizable BAR does not work really on thunderbolt, It is unusable on it.
But let’s have a look on Techpowerup’s tests: In 4k they claim avg FPS of 81% if you use the bandwith, which Thunderbolt 4 can offer. So that’s a absolutely No-Go. But even Oculink with only 4 lanes would loose 8%(with a RTX 4090). But with Oculink 8i (8 lanes) it would loose only 2%, which is basically measurement inaccuracy.
sadly i found no direct tests with thunderbolt 4 vs oculink, but on a bandwith-side techpowerup shows even on avg FPS, that you will loose a lot of perfomance (and as said above, avg fps does not show the real experience). And to say that again: With thunderbolt you have a lot of problems, doesnt matter how good the bandwith is and also important is, that you have to buy one of these proprietary, very expensiv docking-stations.
And also other companies starting offer oculink and NOT Thunderbolt 4 and Framework should do that too (as a expansion-bay). I mean, if i would be able to do it by myself, i would do, but i am not an engineer.
Here are some links about the oculink, which are nice to read:
Yeah, i would also loove to see that or a expansion bay designed by community (so one without battery).
But for me it seems difficult, because when i look on Framework 16" videos, the PCI-E “Slot” is not a slot, it is a plug. So you would need a PCI-E 4.0 x8/x16 female to Oculink 8i female adapter. And i dont find one. Or do someone find one ?
Or would it be maybe possible to use directly a PCI-E 4.0 x8/x16 female to PCI-E 4.0 x16 female riser cable? But i don’t find one of these in PCI-E Gen 4. And i find only 16 lanes to 16 lanes (in pcie gen 3), so i have no idea, if this would be to wide to fit in the expansion-bay because of the 16 lanes wide.
But what i want to say: Maybe it is possible to use directly PCI-E to PCI-E, because when i understand correctly the Framework 16" offers 8 lanes of PCI-E 4.0 directly out of the laptop (there, where the expansion-bay sits). But for sure this would be a worse workaround than Oculink i think, because rasor cable are not that comfortable like an oculink cable i think.
Would also like to know that from people who are completly in this topic. But for me i would wish a 8i version of the one dock, because it is “only” 4i, so 4 lanes, when i am right.
Thinking about exactly the same and for now waiting for the expansion bay deep dive from FW. Then I guess it is about harrassing Optimum Tech or others for files and ModDYI for cables ;). That’s my plan so far… ;).
Yes, don’t worry, just wanted to write that this was code for: I do not have any plan ;).
All I have is a hope that framework would be willing to maybe put an adapter on the market place. I know framework staff mentioned the connection should not be dramatically difficult to make.
And of course there is the hope that the community could come up with a solution, given that framework staff has mentioned it is feasible and should not be too complex.
imo using 2 of the existing 4x occulinks may be a better idea so you can use both to get an 8x link to a special 8x enclosure but you can also use just one of them to interface with existing enclosures made for other devices that use just a 4x link and you might even be able to use the second one for something else depending on how flexible the bifurcation is.
Nah pretty easy. They have design files on their github for the PCB, all we’d need to do is connect up the PCI-E traces to an oculink port (making sure the traces are PCI-E compliant, including exact same length and impedance - not my field of expertiese, still doable).
So then you could go to cheap PCB printing shop (PCBway, jlcpcb, etc) and have it printed for ~20-40$ for 5 or so, and then solder on the oculink port and fan headers (and probably some passive components). Done
I’ll ask a friend to do it once I get my shell if nobody has by then. But I think somebody will, it’s actually easy. Hardest part is drilling holes for the oculink port into the shell
NOTE: GPD G1 and other solutions use the Oculink-8612 port, so only PCI-E Gen 4x4, not x8, like the Oculink-8i. So the custom adapter should probably feature both ports. Then a cheap adapter from oculink 8i to full-size PCI-E (needs external power input for default 75W and some passive components). Existing adapters for Oculink-8612 are so low-volume, they cost ~40€
If it would be “pretty easy”, you would have done it already and would not need to ask a friend
If it would be “pretty easy” no one would ask here for that feature or would have to wait for someone else doing it.
That’s not the meaning of easy. It would be pretty easy, if you can just buy a cable and plug it in. Solder somewhere on a board, drilling holes in the laptop and so on is far away from “pretty easy”.
Yeah, maybe someone makes this, but maybe not and an official one is just better. So that’s why we ask here. Not all are able to design a board by themself.
I mean if i could do it, i would design it. But i have no knowledge of electrical engineering or anything in designing PCBs. That’s not same as playing with Lego
If someone gives me a turtorial, how to do it, i would try it. But that’s not sth you can learn in a 30min turtorial. You would need probably a lot more knowledge of that stuff. It’s for example not like designing sth for a 3d Printer. You can learn that in 30 minutes. But for designing a PCB you need electrical knowledge.
So im open for doing it, but i dont think i’m able to do it.
I dont think this is a good idea. This makes it just more complicated, when you are coming from one PCI-E x8 plug, have to make two oculink 4i slots. Better just make one 8i slot and then you can buy this cable, if you really need a 4i slot or if you really need 2 4i slots:
