# External GPUs **URL:** \ **Category:** Framework Desktop\ **Created:** [September 8, 2026, 12:22pm UTC](https://community.frame.work/t/external-gpus/84700 "2026-09-08T12:22:28Z")\ **Posts on this page:** 17\ **Page:** 1
**Author:** ![Fishd](https://sea1.discourse-cdn.com/flex001/user_avatar/community.frame.work/fishd/32/34059_2.png) [@Fishd](https://community.frame.work/u/Fishd)\ **Post date:** [September 8, 2026, 12:22pm UTC](https://community.frame.work/t/external-gpus/84700/1 "2026-09-08T12:22:28Z")
Adding as a separate topic as it’s more of a question for the community rather than about the product… I see a lot of posts asking about eGPUs or Occulink ports… why? I would have thought those sorts of needs would be best satisfied by a standard mini-ITX board with a PCIE x16 slot rather than running GPUs over PCIE 4.0 x 4, which is going to significantly hamper the bandwidth. I also don’t understand buying a tiny PC to then strap an external case with giant GPU next to it. Is this is “just because I can” or is there something I’m missing? ---
**Author:** ![Camilo\_Reyes](https://sea1.discourse-cdn.com/flex001/user_avatar/community.frame.work/camilo_reyes/32/39421_2.png) [@Camilo\_Reyes](https://community.frame.work/u/Camilo_Reyes)\ **Post date:** [September 8, 2026, 7:24pm UTC](https://community.frame.work/t/external-gpus/84700/2 "2026-09-08T19:24:20Z")
Definitely, it’s a “just because” since the Framework Desktop is very customizable. ---
**Author:** ![Jason\_Sharrad](https://avatars.discourse-cdn.com/v4/letter/j/67e7ee/32.png) [@Jason\_Sharrad](https://community.frame.work/u/Jason_Sharrad)\ **Post date:** [September 10, 2026, 4:29am UTC](https://community.frame.work/t/external-gpus/84700/3 "2026-09-10T04:29:09Z")
Oculink functions simply as a PCIe extension cord, allowing users to add high-bandwidth slots outside of the main system chassis without complex installation. It requires no drivers, setup time, or additional power draw. For demanding graphical workloads, such as running AAA games at 4K resolution with an RTX 5090, Oculink maintains strong performance even when operating at x4 bandwidth. Beyond gaming, the system excels in AI/ML tasks; dedicated GPUs can output tokens orders of magnitude faster than integrated graphics (iGPU). If you use an external GPU solutions connected via USB4—where bandwidth is significantly lower than PCIe Gen 4x4—Oculink offers a substantial advantage. If model size and data handling are managed correctly (keeping all assets within VRAM), the link between eGPU to CPU doesn’t matter so much. Using eGPU’s facilitates complex, multi-stage processing pipelines. For example, one GPU can run a text LLM to generate intricate inputs, such as prompts for video generation (e.g., Minimax H3 prompts, stable diffusion image gens for H3 keyframes), which are then seamlessly passed via USB to the second dedicated eGPU via Oculink for the final rendering and video generation task where the link bandwidth matters more. The type of work that does not work will it fine tuning and training where the bandwidth between CPU, RAM and GPU matters much more. ---
**Author:** ![Fishd](https://sea1.discourse-cdn.com/flex001/user_avatar/community.frame.work/fishd/32/34059_2.png) [@Fishd](https://community.frame.work/u/Fishd)\ **Post date:** [September 10, 2026, 9:03am UTC](https://community.frame.work/t/external-gpus/84700/4 "2026-09-10T09:03:30Z")
I understand the tech, it’s the logic I’m failing to see. ---
**Author:** ![Alan\_Pearce](https://avatars.discourse-cdn.com/v4/letter/a/2bfe46/32.png) [@Alan\_Pearce](https://community.frame.work/u/Alan_Pearce)\ **Post date:** [September 10, 2026, 10:45pm UTC](https://community.frame.work/t/external-gpus/84700/5 "2026-09-10T22:45:33Z")
> [@Fishd](#): > > I understand the tech, it’s the logic I’m failing to see. I suspect there is a certain amount of wanting one machine that will do everything, a high performance gaming machine when at home as well as a portable machine for taking to the office or on trips. Having a laptop satisfies the second part and adding an eGPU satisfies the first part. Using Oculink provides the best link between the two, but USB will suffice where Oculink isn’t available. ---
**Author:** ![Jason\_Sharrad](https://avatars.discourse-cdn.com/v4/letter/j/67e7ee/32.png) [@Jason\_Sharrad](https://community.frame.work/u/Jason_Sharrad)\ **Post date:** [September 11, 2026, 3:24am UTC](https://community.frame.work/t/external-gpus/84700/6 "2026-09-11T03:24:32Z")
Your values are not the same as my values. If you don’t understand why I would want all that on a single machine, then I can’t help you. ---
**Author:** ![Fishd](https://sea1.discourse-cdn.com/flex001/user_avatar/community.frame.work/fishd/32/34059_2.png) [@Fishd](https://community.frame.work/u/Fishd)\ **Post date:** [September 11, 2026, 9:46am UTC](https://community.frame.work/t/external-gpus/84700/7 "2026-09-11T09:46:23Z")
Yeah, the ideal “one size fits all” machine… I’ve been seeking that for a while. Seems I made a mistake thinking about Occulink, I thought it could provide a full x16 lane connection but it looks like it’s limited to x4 anyway, so I guess a perfect fit here… I can only assume Framework didn’t expose it because I believe in some instances, hotplugging it can be a bit iffy. ---
**Author:** ![Agall](https://avatars.discourse-cdn.com/v4/letter/a/a9a28c/32.png) [@Agall](https://community.frame.work/u/Agall)\ **Post date:** [September 13, 2026, 2:00pm UTC](https://community.frame.work/t/external-gpus/84700/8 "2026-09-13T14:00:06Z")
> [@Fishd](#): > > I also don’t understand buying a tiny PC to then strap an external case with giant GPU next to it. I’ve experimented with this, amusingly, the performance (in gaming) was worse with my RX 7900XT via USB4 in a full 40Gb/sec dock which I sorted of expected with how capable the 8060S is. Its a cool ‘because I can’ rather than anything practical, but I have thought about experimenting with using an eGPU for recording/streaming to offload it from the 8060S. Even just recording in OBS drops the framerate by 25%, so even a simple USB4 eGPU as the encoder makes sense, if it even works. Unfortunately the m.2 to occulink adapter I got wouldn’t POST in the machine to try occulink though, maybe because its a PCIe 5.0 model, but I plan to do more testing in the future. Personally I’ve spent a collective 2 years running eGPU setups as my daily driver since 2016, the 8060S being just too capable to be practical outside of potentially offloading specific tasks to. ---
**Author:** ![convexproblem](https://avatars.discourse-cdn.com/v4/letter/c/bc79bd/32.png) [@convexproblem](https://community.frame.work/u/convexproblem)\ **Post date:** [September 13, 2026, 5:05pm UTC](https://community.frame.work/t/external-gpus/84700/9 "2026-09-13T17:05:05Z")
PCIe bandwidth actually isn’t a big deal for a lot of common applications. PCI-e 2.0 x1 eGPU for gaming was only a minor drop in performance compared to a proper x16 desktop setup many years ago when I still played games. I mistakenly thought x4 would be a bottleneck for LLM inference but I realized you only pass a few megabytes per NN layer across so it’s not a bottleneck at all. Even tensor parallel isn’t that slow on the commonly used x8/x8 split. I’m waiting on the correct power connector for my V100 but I’ll be connecting it to my strix halo over a x4 extension cord in a few days. ---
**Author:** ![Ernie\_Hobbs](https://avatars.discourse-cdn.com/v4/letter/e/d07c76/32.png) [@Ernie\_Hobbs](https://community.frame.work/u/Ernie_Hobbs)\ **Post date:** [September 18, 2026, 11:03pm UTC](https://community.frame.work/t/external-gpus/84700/10 "2026-09-18T23:03:52Z")
for me, the size of the case was not a positive- it was definitely a negative, but up to 128GB RAM unified memory was the draw. But I found after I bought it that it was quite as fast and capable as I had hoped for running LLMs, and adding a eGPU (sometimes two of them) was a way to bring it up to what I had hoped I would originally. It is still pretty good, but the plan has migrated over the past year. And there is some of the “just because I can” in it, too. If I had to do it all over again, I would buy just the board and put it on a test bench, so that I can tinker as much as I want without the constraints. A perfect setup would be a strix halo board, 128gb, with three or four extra PCIe x16 lanes available. But that doesn’t exist in this price (which has doubled since I bought mine), so we live with what we have. ---
**Author:** ![Dan\_B](https://avatars.discourse-cdn.com/v4/letter/d/c2a13f/32.png) [@Dan\_B](https://community.frame.work/u/Dan_B)\ **Post date:** [September 26, 2026, 5:21pm UTC](https://community.frame.work/t/external-gpus/84700/11 "2026-09-26T17:21:53Z")
how do you connect the eGPU to the framework desktop? thanks dan ---
**Author:** ![Ernie\_Hobbs](https://avatars.discourse-cdn.com/v4/letter/e/d07c76/32.png) [@Ernie\_Hobbs](https://community.frame.work/u/Ernie_Hobbs)\ **Post date:** [September 26, 2026, 7:26pm UTC](https://community.frame.work/t/external-gpus/84700/12 "2026-09-26T19:26:51Z")
I have two of them. One is Oculink (an adapter using the m.2 slot), and the second is connected via USB4. I’m happy to share the adapters/eGPUs I bought- just let me know if you’d find them useful. ---
**Author:** ![Dan\_B](https://avatars.discourse-cdn.com/v4/letter/d/c2a13f/32.png) [@Dan\_B](https://community.frame.work/u/Dan_B)\ **Post date:** [September 27, 2026, 9:17am UTC](https://community.frame.work/t/external-gpus/84700/13 "2026-09-27T09:17:29Z")
please share the adapters, that would be great. I assume you cannot have (2) M.2 drives if you are using the slot for Oculink Which USB is the USB4? thanks dan ---
**Author:** ![convexproblem](https://avatars.discourse-cdn.com/v4/letter/c/bc79bd/32.png) [@convexproblem](https://community.frame.work/u/convexproblem)\ **Post date:** [September 27, 2026, 12:42pm UTC](https://community.frame.work/t/external-gpus/84700/14 "2026-09-27T12:42:33Z")
I used a M.2 A-E key to NVME adapter for my second drive (it’s a slower drive set up for daily backup, not mirroring). My PCIe x4 slot is broken for some reason but if you remove the metal bracket a lot of adapter cards are shorter than the Framework heatsink. I’m using this M.2 nvme to PCI-e x4 adapter, it’s a 4.0 rated 75cm cable that looks like Oculink but apparently isnt: https:/e.tb.cn/h.8DhgUyi5sug68Rx Apparently it’ll actually successfully connect at 5.0 too but these strix halos are only 4.0. AFAIK PCIe 4.0 unlike 5.0 isn’t very demanding on the cable, so anything that says 4.0 and has robust 12V should be ok. The sellers who make ribbon cable risers with SATA power plugs claim those do OK too but technically the 12V contacts are only rated for 4.5A so I’m using a 4 pin molex to 6 pin adapter on mine. ---
**Author:** ![Agall](https://avatars.discourse-cdn.com/v4/letter/a/a9a28c/32.png) [@Agall](https://community.frame.work/u/Agall)\ **Post date:** [October 1, 2026, 7:01pm UTC](https://community.frame.work/t/external-gpus/84700/16 "2026-10-01T19:01:45Z")
I’ve experimented with both USB4 and oculink using a couple docks for USB4. The oculink adapter I’m using that works is this one: [Amazon.com: OwlTree M.2 NVME Pcie to OCuLink SFF-8612 Adapter, PCIe 4.0 x4 64Gbps to OCuLink SFF-8611 4i Host Adapter for eGPU GPU and M.2 NGFF SSD 2230 2242 2260 2280 (5.9inch) : Electronics](https://www.amazon.com/dp/B0GFN45LW6?ref=ppx_yo2ov_dt_b_fed_asin_title) There’s another one I originally bought that didn’t work, likely because it was PCIe 5.0 4x, but the one I have linked has been working fine. I have my own modified version of Noctua’s side panel that I added an oculink tool-less mount for. I’ve had issues uploading this to [printables.com](http://printables.com) since its licensed under Noctua, but I might try again at some point. ![IMG_1879](https://us1.discourse-cdn.com/flex001/uploads/framework3/original/3X/9/d/9d57bc6528143cf121066447d9584198118cf03b.jpeg) The thermal performance without the side panel is pretty massive to where using the front side board m.2 slot isn’t worth it if you’re not using a modded side panel. What’s likely better to do if you don’t have a 3d printer is to use the rear m.2 slot without the panel installed. I’ve had no issues with thermal performance having my main m.2 drive installed on the rear m.2 slot with the included heatsink. The only problem I’ve run into with oculink eGPUs is the machine won’t POST on the eGPU, so I’ve had to make sure my single display setup is plugged into the iGPU, then swap it to the eGPU once its POSTed. I’m running SteamOS on my Max+ 395 64GB version. ---
**Author:** ![Jason\_Sharrad](https://avatars.discourse-cdn.com/v4/letter/j/67e7ee/32.png) [@Jason\_Sharrad](https://community.frame.work/u/Jason_Sharrad)\ **Post date:** [October 2, 2026, 2:07am UTC](https://community.frame.work/t/external-gpus/84700/17 "2026-10-02T02:07:20Z")
That’s a very nice cut in your panel, much better than i did. ---
**Author:** ![Agall](https://avatars.discourse-cdn.com/v4/letter/a/a9a28c/32.png) [@Agall](https://community.frame.work/u/Agall)\ **Post date:** [October 2, 2026, 2:28am UTC](https://community.frame.work/t/external-gpus/84700/18 "2026-10-02T02:28:11Z")
Not cut, I designed a snap in mount for the board of the m.2 to oculink adapter. I uploaded the model for printing: [https://www.printables.com/model/1862512-framework-desktop-oculink-noctua-side-panel-remix](https://www.printables.com/model/1862512-framework-desktop-oculink-noctua-side-panel-remix)