GPD G1 + Minisforum UM790 (and a chat about Oculink and Thunderbolt Next) 

"you wouldn't download a level one techs wallpaper"

The best part of industry standards is there are *so many* to pick & choose from!

I like the blown ambiental audio and day room lighting... authentic computer janitor

Fantastic video Wendell. GPD has really hung in there over the years - I still have a GPD Win 2 I use a few times a year...I really should turn that into some kind of crazy Frankenstein server or something in the spirit of L1T.

I think more influencers/tech reviewers should be critical about lack of BIOS updates on these miniPCs... having none at all seems to make product obsolete as soon as some bad implementation issue or CVE comes around. That is why tinyminimicros from big brands have at least some sort of time limited support for that... sadly it gets ignored by reviewers

An Oculink to MXL adapter would be interesting. But what I really want is an affordable PCIe to SFP converter that would provide proper PCIe over fibre without exotic drivers.

thanks for bringing such interesting tech content. I would never even think to search for these products.

Level1 Techs is leading the way on "painting outside the lines" 👍 Looking fwd to experimenting with OcuLink and MCIO . . . some day eventually. Thank you. Kindest regards, neighbours and friends.

"Mini Cool Edge" sounds like a name Homestar Runner would come up with.

Speed is not for drives only. I like to see connectivity like that.

Thanks for this detailed explanation

I feel very well informed and entertained. Guess thats what its all about. Thx a lot

Wife and I both switched to using the UM790 + GPD G1 as our main rigs when we moved to California in December. Honestly impressive. There're some hiccups here or there with the 7600M XT not having that 12gb - 16gb of VRAM but pretty much every game we throw at our setups works flawlessly with her running LMDE6 and me running Fedora.

I love this, so much yet to explore

I wanna see you DIY your own egpu solution, you mad scientist.

I am very interested in another Minisforum review - this time of the BD790i, which has the 7945HX with the extra vCache. I have an old Ryzen 2700X gaming build in a mini-ITX cache, which is very long in the tooth at this point. Working in that case is a real bear of job, but dropping in that BD790i mobo into that case would save me a ton of headaches and grief (like spending an hour trying to reach near impossible to get to fan headers).

Wendell, you are something special

I would be very satisfied, if Thunderbolt could give me the option to only run PCIe over that USB-C (Laptop to eGPU) and the dock specific part over that USB 2.0 separate stream that every USB-C cable should include. Keyboard, mouse and audio don't need more than USB 2.0 and if I need any high-speed connectivity I would directly connect it to the laptop. What wouldn't work is any display signal over that cable in my case and i would be fine with that too. The internal display could still run with the iGPU. Of all that might be very complex but it would be really nice. Even better when the system could switch between running on the iGPU with USB4/TB working normally and then that mode i explained above. One can dream.

Thanks a lot for these insights into Oculink, PCI & Co.! Not that I really understood all of it, but nevertheless... :) As an aside: are you trying to interface that HP 100LX to the Minisforum PC? Could I reactivate my old 95LX in some way??

that's really cool! i hope i can buy something like this some day.

It's a very refined object, but if I think about how much a simple, banal port costs that has a real 40Gbit of bandwidth to move a few files quickly (where the filesystems allow it), it's actually not bad at all. It's like having the graphic card separate from the machine plus a top-of-the-line dock included in the price. It is a product that needs to be understood, which is not suitable for everyone, but which makes a lot of sense in my opinion, for many many people, especially if you consider that even laptops, when the graphics card really works, still want a power socket near. I like it. Thank You!

Not really a fan of the G1, unless you need the portability the value is pretty poor. For the same price you could get a 7800XT. Oculink tho, Mother of God, what a game changer. I've been running a GPD Win 4 with an external 6700XT for some months and the thing is an actual desktop replacement.

Can you get longer Oculink cables? Maybe fibre-optic? Do you see the technology developing such that GPUs become accessories that plug into monitors rather than PCs?

Jesus, you're on GN so much I straight didn't realize I wasn't subbed 🤣

Love to see more expansion options. Only recently started watching your channel, but I will have to hunt down and see if you did anything related to the ASUS ROG XG connector. I really think that's the connector to beat. A combo TB/USB + PCIe with 100W or greater power delivery would be ideal to slim down your cables from 4+ to just 2 (1 power and 1 connecting the dock). Can't wait for more developments in the space.

Awesome video, I have the Win Mini and have been eyeing this device for awhile. However, I wouldn't be using it with an enternal display. I have a 3080Ti for that. My use case would be sitting on the couch at night wile gaming on the internal display. Is this doable or even worth it or would it be too uncomfortable in this use case?

@level1techs Wendell this was a superb write up! I picked up a GPD G1 a few months ago for use with my ThinkBook 14+ via oculink and couldn’t ever get it to handshake properly. No issue via usb4 but the machine never detected the PCIe device. I’d love to send you some system info and get your opinion on the communication failure.

Hey Wendell i recently did a cool setup with a minisforum setup. MiniPC+hypervisor with dual lans and TB3/USB4 drive bay. I have it running a server and router. works great! lol

currently using oculink egpu as my main system with beelink mini pc ive used a bunch of 5 different oculink docks and the only one that works is osmeta g38 which cable isn't detachable, it holds the signal quite well with pcie 4.0

Hey Wendell, can you do a deep dive on PCIe bottlenecking? It would make an excellent companion for your videos on creative ways to get access to PCIe lanes and every other analysis I've seen of bottlenecking assumes that GPUs are the only PCIe devices and that gaming is the only use for GPUs (that, or crypto where the bandwidth requirements are so low that it doesn't really matter). For example, do AI workloads require as much PCIe bandwidth as gaming, and under what circumstances will those workloads experience bottlenecks?

seems like that occulink adapter would certainly work with the couple of Gen3 M.2 2280 daughter board options for the RPi5 - guess its just down to drivers and time needed for someone to "give it a whirl" (I think there are still some issues with BAR I/O for GPU's on RPi5 like there were for RPi4) - As for RISCV, just need a board with usable Gen3 / Gen4 PCIe connectivity - actually there are more than a few 3588 boards with suitable PCIe connections (and less BAR I/O issues) that might be worth trying as well ..

There’s a ton of Oculink to pcie adapters you can just plug a desktop into

I’ve been using my eGPU for AI, so throughput hasn’t been much of a concern, but I appreciate the explanation.

What! An American using the worlds measuring system? What is going on. Sounds like a 3’3” cable to me. lol. Love this guy.

8 gen 4 lanes is the sweet spot for now I think. That would be able to run all but the highest end of cards at full speed. 4 gen 5 will only be good on next gen cards.

Pretty sure I destroyed my oculink port on either the GPD G1 or the GPD Win max 2 (7840U/64GB version) because I hot plugged it the first day I got it without realizing it would destroy it. I have never been able to get it to work and I haven't had time to test the M.2 port with the adapter to confirm if this is the case. Only works over USB 4 in all my tests on Windows10/11. Sending my machine back to GPD set me back $200 through DHL for a free repair when the HDMI died on my other GPD Win max 2 (2022 6800U/32GB version) was a miserable experience but GPD does honor the 1 year warranty at least. It's great to see how awesome everyone's experience CAN be but mine has been... expensive and under-delivers.

Haha, love the shout out to Jeff Geerling - the only person crazy enough to recompile the Linux kernel 15 times to try get an esoteric CPU and GPU combination working. Another good shout out is @TakiUdon who manages to find Vulkan-enabled GPU drivers for SBCs and successfully runs Windows games on ARM Linux using Proton + FEX. Would be fun to see someone try to pair this e-gpu with an ARM system and test some Windows games. I wonder if Thunderbolt E-GPUs are supported on Asahi so we can try using a Mac Mini and side-step the lack of GPU drivers 🤔 Exciting times in the CPU space, can't wait to see more ARM devices (especially Linux compatible ones) hit the market!

Technology has come a long long long ways

What's that arm on your Gigabyte monitor? It's amazing! Essentially you can lay it flat!

Where are the Oculink external drive enclosures? I can’t find anyone selling them anywhere.

Great topic

Wendell, Have you tried VGPU with an oculink?

It is worth noting that (if my understanding is correct, it may not be. Please correct if so): USB4 is technically capable of pcie4x4. But the motherboard side controllers are not. Although it has been proven in testing that USB4 does already exceed the pcie bandwidth of TB3/4 - but this is more a saturation of the pcie3 bus. Not jumping up to pcie4. Anyway, long story short - as the motherboard side controllers progress, the USB4 connection of the GPD GPU may well come close to the OCULink connector... You know, once the connector breaks 🙃

I cant wait for mine

I would LOVE to have this for my work laptop since the iGPU just makes the fans scream all day when I'm using two monitors + the laptop screen but the price tag just isn't right. A full RX 7600 is only $260. If they could get this down to $300-350 then the investment might be worth it for the long term.

Are you sure about Type C ports supporting up to 8 high speed pairs? I understand you said it has not been implemented, but I'm not seeing enough pins to support that unless some of the auxiliary pins are optional. 24 pin total (4) for the two TC pair (4) for the two RX pair (4) GND (4) VBUS (4) for reversible USB 2.0 (2) for SBU (2) for VCONN/CC1 I suppose the USB 2.0 pairs could be reworked (who wants USB 2 anymore? :) ) and the SBU pair is open for alt mode, but that would only open up 7 pair. I would think reworking any of the power pins would be a recipe for disaster, but I could be wrong.

I don't get one thing. eGPU community recently showcased an m.2 to usb+oculink interface for the ROG Ally. What I don't get is how it is claimed to be faster than a normal m.2 to pcie riser? Even if both are on the same gen4 wouldn't 4lanes of m.2 bottleneck them both, equally? I think it was touched upon this video but I still need some more simplification to understand.

I running my GPD G1 with my zenbook flip (i5 12500H) via TB4. It does really well and is great for "on the road". I would say it is comparable to my TB3 RTX 3070 eGPU setup (in a razor core-x). I have also found some software (Lossless Scaling cf steam) to also get upscaling and frame gen, and can offload some of that onto the iGPU. This is good as Adrenalin drivers don't do well with hybrid GPU setups I find (as in no AFMF etc..) . I'm playing older games like FFXIV just great and also modern ones like Cyberpunk. In future maybe a TB5 machine if they make a TB5 to Oculink adaptor :)

btw you don't even need the full bandwidth for even the most modern GPUs. I get like 90% of the performance out of my RX6700XT despite running on a 16x PCIe 2.0 interface. It really only matters during times where you are loading assets from storage, and even then my SATA SSD is the bottleneck, not the PCIe link, and with 12GB of VRAM this GPU basically has enough memory to save the entire texture and lighting information for the whole game in most games I play. Sure I probably have worse latency than I would have on other systems or some other slight hiccups, but nothing I really ever notice. Now if this was a machine learning situation with RAM cacheing I would run into issues with this, but the way it is now, it's alright.

I tried a thunderbolt 3,4 GPU dock for my Minisforum UM 790 pro. It cost me £120 plus £85 for a Power supply. I already had a RTX 2080. Although it was a huge improvement I WANTED MORE🤣so I though I would try an Oculink setup(which was cheaper than the thunderbolt setup) In the meantime I took my RX 6900 XT from my gaming rig and did a little comparison using Unigine Heaven and it blew me away(do they still say that 🤣) Anyway here are the results if your interested: Minisforum um 790 pro(780m) with graphics set to 6GB in Bios : - FPS=77.8, Score 1960 Nvidia RTX 2080 with internal graphics disabled : - FPS=298.7, Score 7524 AMD RX 6900XT with internal graphics disabled : - FPS=391.4, Score 9859. I Thought the 790 Pro was quick but the above results brought me down to earth😃. I suppose I should run both cards through my gaming rig and see just how much faster Pci 16x is than pci 4x(Oculink through the M2 Nvme).

WHAT is that widescreen device under the monitor? That's exactly the form factor I want for a sort of minipc or cyberpunk layout...@4:49 ish

Would you mind to share the actual cable connecting on your device? I got this exact same setup, inspired by ETA prime tests, and got really disapointed because the M2.Ocullink connector does not go well inside the Minisforum UM790 Pro. Now I am using tb4. Moreover, I am getting some wierd driver issues, because my system goes dark when I am playing really graphic intensive games. Using on PopOS, did not try on windows.

Thunderbolt is cool. 7600m doesn't get better than it actually is. Oculink is that nice thing that PC users never had. Wish half PCIe slots were ripped out from PCs and replaced with such connectors,while also being supported by market for regular people. All those PCIe risers could be ditched,things could be smaller and more compact, pretier and much more practical.

Cool, I believe I had read at the inception of TB4 that even though the bandwidth max was the same as TB3, in real life applications, TB3 was inferior to TB4 because TB4 had the same throughput but biderectionally simultaneously, effectively twice the bandwidth.

GPD Win Max 2 2023 external OCuLink connector by default is forced to PCIe 3 as of now due to instability problems GPD got with the very specific lanes used for the connector - if you force PCIe4 in the BIOS for it it disables CPU boost as they could not handle PCIe4 and CPU boost at the same time and AFAIK were working with AMD to solve this. The 2230 M.2 PCIe lanes are fine and if you add OCuLink connector there it can do PCIe4 and CPU boost at the same time.

Oculink has peaked my interest (and bandwidth).

It's really nice to see a company create this. So tired of the regular gpu market only cater to desktops with giant graphics card. But price. I understand it's not mass market product, but the price is only $400 less than getting a full fledged Ryzen 9 7845hx and 7600M XT gpu Alienware M16 from Dell. I would also like to see something similar made for storage. Specifically, m.2. or maybe even 2.5" ssd. I understand it would need a switch and not be full speed, but that is fine.

of the minisforum elitemini UM 780 xtx, the geekom A7, and the beelink gtr7 which would you recommend for someone wanting to game and do some light data science?

Hello! I'm not that knowledgeable on this, but i would like to ask about the temperature? Both the cpu and the egpu? Does it spike? Or does it stay at a normal range? Thank you

I'm still on Thunderbolt 1 with my Thinkpad T430s

Am I missing something? The UM790 Pro doesn't have Oculink. You'd have to use one of the "M.2 Cheat codes" cards - or the type from the UM780 XTX. Is this an UM780 XTX. It's angled like it is using the UM780 XTX stand.

4:30 "But let's not be super pedantic" - But that's the whole point of L1T. That's WHY I WATCH L1T. 😜

Now if only we could get better GPU's in the oculink docks, oculink is very exciting actually having the bandwidth needed for GPUs, but so far it seems like everyone is just sticking the lower end mobile cards like 7600M XT, it would be far more interesting if they had a more capable desktop card like a 7800XT or 7900XT/XTX.

I was thinking of repurposing aw graphics amplifier but 40 series are not supported.

which is a better idea: minisforum um790 pro with g1 via tb3 or um780 xtx with g1 via oculink?

What's with the audio? Sometimes sounds underwatery or like artifacty. Reminds me of the effect I get when applying too much noise reduction in audacity to try and get rid of a fan or other background noise.

shout out to Jeff !

Modern efficiency in computer chips is wild. I currently have a big, powerful desktop, but I think in a few years I'll probably migrate to a handheld with an eGPU. With the 7840u, he current CPU speeds are plenty for me, I can get 64GB of RAM in the GPD devices, which is more than enough for all my productivity work, and the integrated GPU is plenty in handheld mode (I'm pleased even with my Steam Deck OLED when it comes to handheld gaming). The one thing remaining is improving the eGPU speeds. I have a 3080, so the G1 would be a downgrade for me. In ~3-4 years though, I expect there to be an eGPU that is a significant upgrade to my current GPU.

Probably the best video to date on the topic...well done

TB3 is veeeery outdated. I use it for plugging in my work laptop, it's super handy to have charging, USB and display output in one connector BUT I can only have USB2.0 or I have not enough bandwith for 3440x1440p with 120Hz (luckily "dock", which is built in my display alows to redirect lanes, so I can have high refrashe rate or usb3.0 if needed) and chargning is limited to 100W. TB5 is needed now and for sure 120/40 or 80/80 is not enough for future.

I wonder what happens when I connect oculink to one device and usb 4 to another.

The power connector is on the wrong side! Why? It should be on the oculink side as it’s an input, where as the io is all output! Group your connection! This is why people have crazy cable monsters

Hey can you please thest if Daisy Chaining this to a TB dock would work? Like this: Monitors --> GPU Dock --> TB4 Dock --> Notebook? I know it's ridiculous but I just want to know. I don't have oculink devices jet

Is it a good option for audiovisual editing with Adobe applications?

Better is using an amd 7600 xt using a occulink to pcie 16x adapter

I hope they will do the Thunderbolt 5 fast , its so usfull for Egpu

I really want an add-your-own-card Oculink enclosure. Took a punt on the OneXPlayer X1 and I have a couple of graphics cards spare. Anyone got any recommendations?

I use my 780xt with a 2080ti in a dock I made from an amazon oculink kit, s 650w psi and à socket metsl box (was mastercraft lol). It noe live as my "modular all in one computer " on the tv arm in my truck for night gaming or waiting to get loader gaming. Love that thing, I had to lock the gpu at 60% power for my truck batteries sake buuuut I got 90 to 120fps in gaming and I just bring the cpu with me on week-ends for ipdates and such. Love oculink! I'm sadly running windows, nvidis driver made troubles on Linux. Wish I had an amd card but the 2080ti was the one I had soooo.. meh

Why is thunderbolt so expensive? It seems like you can't get into a thunderbolt motherboard unless you really shell out.

I have UM780XTX + GPD G1. Manufacturer promise that i can start system only with oculink but it doesnt. Only if USB4 connected. I can't name GPD G1 silent, cause its mainly stands on same table wtih monitor(near to user) and has turbine mode cooler, when it goes to "full power" mode(in gaming, for example) noise become very disturbing. Even if you are use headphones. P.S. You can hear this whistling noise in video also.

I kinda hope that GPD would get drunk enough to upgrade to the 8 lane version of oculink.

problem is no matter the bandwidth you can force through a connector if the if the internal and/or external gpu doesn't have the ram capacity how much you the gpus will need to 1440p, 4k or 8k we can't even saturate it 40gbps or even 80gbps 120gbps is nothing bu a myth aswe don't have gpu ramin play to support it

Why doesn't Thunderbolt/USB4 allow "lane bonding"? If the controller was wired to the host with a wide enough bus, and the endpoint (i.e. eGPU dock) had such a "wide" controller aswell, one could connect to an external device with more bandwidth simply by using two cables. It's super frustrating that the development of external ultra high speed IO is being delayed and doesn't see much adoption in the consumer segment. Btw., ASUS made some ROG XG Frankenstein solution where they combine a TB port with an OCuLink port in a "single connector" - this allows using PCIe 3.0x8, although I think they "just" reroute the 4 PCIe lanes going to the TB controller to the OCuLink connector on the fly, in order to reach full speed and then use the former as a "dumb" USB side channel.

I LOVE the idea, but the price makes it not really a valid option.

Thanks for this video. I agree with you Thunderbolt 5 is already obsolete. Intel don't want to accept it. By the time Intel get Thunderbolt bandwidth to where it should be to support extrenal GPUs (eGPUs), integrated GPUS (iGPUs) will be adequate enough to make eGPUs pointless. Laptop discrete GPUs (dGPUs) are already adequate enough to make eGPUs pointless. Intel should have launched Meteor Lake (Core Ultra) with integrated native support for Thunderbolt 5. If the MSI Claw had Thunderbolt 5 it would be a huge advantage to what AMD has to offer. Other than external graphics and docking, Thunderbolt Share is something that should be already in use for multiplayer applications.

You're talking about TB5 being obsolete as soon as it comes out abd I'm still hoping for TB3 and 4 stuff and my brand new $4,000 work laptop with a 12850HX and an RTX A6000 doesn't have TB4. Lol

will this pc handle gray zone warfare good

I'm kind of interested in the other way around; accessing the iGPU on a 780M based system. Or, rather, bifurcating an x16 PCIe lane into four x4 occulink connectors and using four 780M iGPUs in parallel in training LLMs. 96GB of RAM in each of the systems isn't too expensive (strictly speaking this would have to be an 8700G desktop, not a 7940HS mini PC) when compared to the cost of even a 48GB professional GPU, but I'm not even remotely sure how one would configure the client systems with the iGPUs to afford the iGPU as much memory access as possible. Regardless, it would be a very fun project, and some back of the napkin math shows you could actually hit better bandwidth per dollar than a lot of Apple setups. Plus, I think there's some custom GPU kernel fuse operations you could do with Transformer networks to run a GPU kernel, return the layer to the CPU, do some sort of evaluation, and then return to the GPU for another fused kernel to minimize memory access, but that'd start getting into some real CUDA engineering.

"You're very clever, young man, very clever," said the old lady. "But it's PCIe all the way down"

Ya ya ya, but do you use those speakers in the background? I'm being facetious. Great intel... terrible room bloom in the audio as usual. Good presentation Please add a different color shirt for the rooms you are in. Go red or blue or large checkered or even stripes or a shirt with your logo in some kind of metallic color band mix. Something. Time to step it up yet again. You were mentioning obsolescence, welp, tic-toc Mr. tech, no time to dilly dally.

Great Idea that thing. Unfortunately, AMDs mobile GPU lineup consists only four nearly identical mediocre solutions with only 8GB VRam, which was at the edge of being not enough in 2023 and even more today. For that money, I would take the 7900m but not a 7600XTm (which is nearly non-existent, even these exclusive Alienware Laptops are not really available, where never in Germany where I live). But, there are adapters which can provide a PCI-E slot in which one can add every desktop GPU. Would be interesting to test how bad only 4 lanes are for a 7900-Class Desktop card. I would suggest taking a 7900GRE (260 Watts) and bring it down to around 200 Watts, which is the maximum the 7900m is rated with power-shift (or how its called). Should be the better solutions and these 7900GRE are comparatively cheep. I'm tempted. I have a beelink GTRpro and its Ryzen 7940hs should have enough power to feed the 7900GRE. And it would cost the same a top-of-the-line power-supply included.

I wonder if Thunderbolt 5 > Oculink will be possible.

Doom Eternal sure is performant.

What is the justification for Thunderbolt if we could just Oculink-PCIE everything that requires high speeds?

It's a damn shame Framework didn't go with OcuLink on their laptops instead of the janky GPU module on only the 16.

We far of having 1 million insertions in USB like 80million clicks on keyboard and mouse

it's impressive I'll admit but given the price you're still better off with a traditional desktop

anyone the model of the small laptop wendell was using 21.22 ?

you can create a save box with 4tb there you start a live co-control program like a steam and the same on your system then you release the system there and back and it starts via console and streams through your copy in 60 Hz with a console You can set up a recording program as a stream with good transmissions, 1x1 copy with taxes and then you turn on your program normally in 4k and steam the console stretched out and the edge invisible and start the 1440p application over it to 1080p or 720p for more Hz performance Everything about a monitor that you can use to assign complex data Internet and system a turns graphics via consoles and programs the transfer and processes no time for something like that you can basically accelerate and transfer applications correctly with it with the best for more FPS maximum resolution applications and full graphics. Support everything possible if I could. Worth it, all of you unemployed, I have to go to work now, sorry, good luck

The Valve index headset also uses the oculink connector