"OGRE-250" .67L OCuLink eGpu module, using HDPlex GAN 250w

[SFFMunkee]

@msystems did you have any issues with slot-powered GPUs connecting via OCuLink?

For some reason my A2000 refuses to clock above 1200MHz and seems to be power-limited at about “25W” (actually 50W). No settings seem to change this.

I’ve tried two host machines (X300 and Thinkpad), two M.2 to OCuLink adapters, two power supplies. It worked fine previously on either machine/PSU using an M.2-PCIe direct riser, so I can only assume it’s the powered riser or possibly the cable.

As I don’t (yet) have alternate cable or riser I figured I’d ask if you’ve experienced anything similar and how you got around it. (What was your experience with the Intel DeskMini boards, btw?)

 

[msystems]

@msystems did you have any issues with slot-powered GPUs connecting via OCuLink?

For some reason my A2000 refuses to clock above 1200MHz and seems to be power-limited at about “25W” (actually 50W). No settings seem to change this.

I’ve tried two host machines (X300 and Thinkpad), two M.2 to OCuLink adapters, two power supplies. It worked fine previously on either machine/PSU using an M.2-PCIe direct riser, so I can only assume it’s the powered riser or possibly the cable.

As I don’t (yet) have alternate cable or riser I figured I’d ask if you’ve experienced anything similar and how you got around it. (What was your experience with the Intel DeskMini boards, btw?)

For me it operates like m.2 except the link stability issues.

Which riser are you using? When you say it's power limited, is it crashing or its just acting power limited?

 

[SFFMunkee]

For me it operates like m.2 except the link stability issues.

Which riser are you using? When you say it's power limited, is it crashing or its just acting power limited?

I’m using this one https://a.aliexpress.com/_mPXBdHI

After some further testing, I’ve validated that it works 100% with the 6600XT so I’m assuming it’s a slot-power limitation.

That is, presumably it’s only wired for PCIe x4 INCLUDING 12V/GND. Whereas I’m assuming the powered M.2 to PCIe x16 have only PCIe x4 signalling but still have the full set of 12V/GND pins connected.

 

[msystems]

Hmm, that board looks very similar to the SATA version one except it takes 24 pin.

I wonder if it was just lazy design and although the input format is 24-pin ATX, it doesn't actually have better ability to handle the current (same thin traces and schematic except connector changed). Does it get warm to the touch?

If you have a fine tip solder iron and are feeling brave you can add more +12/gnd leads directly from the little pins on the pci slot to the 24-pin board connection.

 

[SFFMunkee]

Hmm, that board looks very similar to the SATA version one except it takes 24 pin.

I wonder if it was just lazy design and although the input format is 24-pin ATX, it doesn't actually have better ability to handle the current (same thin traces and schematic except connector changed). Does it get warm to the touch?

If you have a fine tip solder iron and are feeling brave you can add more +12/gnd leads directly from the little pins on the pci slot to the 24-pin board connection.

I figured I'd test out a few other units as well, as the form factor actually doesn't work for what I wanted.

What I want is something the size of the SATA-connector model, but with a 4-pin mini-fit jr (i.e. ATX12V) for 12V/GND supply.

The end plan is to either:
- combine DeskMini X300 board + GPU (A2000 or 6600XT) in the least possible volume (aiming for <2.5L with A2000 or <4L with 6600XT)
or
- build a tiny brickless eGPU that doesn't require a relatively expensive HDPlex supply, that I can easily connect to my laptop, or DeskMini or any other PC with M.2 or PCIe connectivity (assuming an OCuLink adapter is installed already).

 

[SFFMunkee]

Does the PCIe signalling require any power from the SATA connector? Or would it be possible to use it purely for the PCIe signal, and have a powered riser attached...?

The other option is maybe an OCuLink to M.2 with a powered M.2 to PCIe x16...? Hmmm

i.e. an eGPU enclosure with FlexATX/DC-DC/12v input/whatever and an OCuLink connector on the back
Internally it would convert the OCuLink to M.2 then to PCIe x16 with the power supply feeding that riser and any supplementary PCIe power connectors if required by the GPU.

 

[msystems]

Its just data and 3.3v, same as m.2 is. You add external 12v on the riser. It looks like some of the risers handle the 3.3v differently as to whether its just passed through. On adt link m.2 to x16 it only needs 12v. The ali oculink board with sata tries to create 3.3v from 5v I think. That's what the caps and the linear regulator is doing on the left side i think.


I think OSMeta makes a oculink to x16 riser with 6pin power input

 

[SFFMunkee]

Its just data and 3.3v, same as m.2 is. You add external 12v on the riser. It looks like some of the risers handle the 3.3v differently as to whether its just passed through. On adt link m.2 to x16 it only needs 12v. The ali oculink board with sata tries to create 3.3v from 5v I think. That's what the caps and the linear regulator is doing on the left side i think.


I think OSMeta makes a oculink to x16 riser with 6pin power input

Yeah Revo posted the OSMeta riser in another thread

 

[SFFMunkee]

So, good news - with a 'frankenstein' of risers (M.2 to OCuLink -> OCuLink to PCIe x16 -> PCIe x16 extension -> PCIe x16 to dual x8 with 4-pin power) I managed to get the A2000 to pull full power!
This means that I'm sure I can simplify the design down to a basic OCuLink to PCIe x16 with 4-pin power and it should work perfectly. Pity there's no off-the-shelf product for this yet

::EDIT:: Let me retract that, it seems the riser only works with the ATX 24-pin connected. If I supply only the 12V to the slot directly it doesn’t work. So it seems the riser actually uses the 5V or 3.3V from the 24-pin connector. Easily solvable but wanted to correct my previous statement !

::EDIT:: I'm thinking the cleanest and smallest method will be:
[PC: M.2 to OCuLink ] <- OCuLink Cable -> [ eGPU: OCuLink to M.2 -> M.2 to PCIe x16 +12V power ]

 

[msystems]

For some reason I don't see the shorty versions of the OSmeta risers anymore with 6-pin input, they all seem to use 24 pin. But anyway if you get your hands on one they do have the benefit of adding a switch on the pcb to choose whether to power the card automatically on signal. I have experienced some strange behavior with this not always working, but fortunately there is a hard "on/off" switch and the sync "on/off" switch.

Superbuy - The Best Taobao Agent Help You Shop,Shipping From China

The best way Shop From China. Browse over 1 billion products from China's shopping platforms(taobao.com,1688.com,tmall.com). Help customers buy chinese products and shipping to gobal. Free purchase, quality inspection, photography and 90-day free storage.

www.superbuy.com

 

[SFFMunkee]

For some reason I don't see the shorty versions of the OSmeta risers anymore with 6-pin input, they all seem to use 24 pin. But anyway if you get your hands on one they do have the benefit of adding a switch on the pcb to choose whether to power the card automatically on signal. I have experienced some strange behavior with this not always working, but fortunately there is a hard "on/off" switch and the sync "on/off" switch.

Superbuy - The Best Taobao Agent Help You Shop,Shipping From China

The best way Shop From China. Browse over 1 billion products from China's shopping platforms(taobao.com,1688.com,tmall.com). Help customers buy chinese products and shipping to gobal. Free purchase, quality inspection, photography and 90-day free storage.

www.superbuy.com

Yeah I was looking last night on TaoBao and could only see the 24-pin versions

 

[CC Ricers]

Looks like I might jump on the OCuLink train with an eGPU after I learned to like the LattePanda 3 Delta again. It's no chump for less demanding games, but it also has a M.2 slot available for PCIe card expandability.

I have a Radeon WX 4100 that has proven to work with the LattePanda using SATA power with a traditional PCIE riser adapter. So would it be able to be plug and play with OCuLink as well? I like the thinner cable which would make it pretty nice to use the LattePanda in a standalone portable device and add the option to dock it with a eGPU for more power as a desktop.

 

[msystems]

In general if GPU works over m.2 it should also work over Oculink, as both are just raw PCI lanes. But its not 100% so probably can't say for sure until its tested with the hardware. Use a short cable if possible if trying to go for gen 4.0. For gen 3.0 it doesn't seem to matter as much, its easy to achieve gen 3.0 with .5 meter cable.

 

[mattafaak]

Thanks! Since there's a little interest, I'll go ahead and design an aluminum version and get a quote for 5 units

I'm buying the parts for this as I await a GPD Win Max 2. Definitely in for an aluminum enclosure with 40mm fan hole.

 

[Justalurker]

With the release of the 500w GAN im looking to update the design to support that psu. The 500w gan has an integrated fan which remedies the cooling situation, so it won't need a metal housing or separate fan. It also will be able to handle the most powerful cards, and the cost is only $30 more than the 250w gan. Its nearly perfect for an egpu.

You can read about the 500w gan here

HDPLEX 500W GaN AIO ATX Power Supply

Free Shipping Worldwide Stock Status USA : In Stock Germany : In stock Asia : In Stock UK : In Stock

hdplex.com

Hey msystems! any updates on the 500w design? I finished my DeskNano build a couple months back and am loving it. An Ogre-500 would be amazing for futureproofing. Thanks!

 

[SFFMunkee]

Hey msystems! any updates on the 500w design? I finished my DeskNano build a couple months back and am loving it. An Ogre-500 would be amazing for futureproofing. Thanks!

SECONDED!

 

[msystems]

I have not had time to work on it unfortunately. I think that OneDock is a pretty good fit for when portability is required though.

For stationary eGpu installations, I have thought of a better solution using OSMeta GK01 which I have the hardware for but no time to finish the design. Basically would be four different enclosures of slightly sizes and gpu length, with flexible psu support for HDplex, Flex, SF and ATX enclosures. Then the user can optimize for card size.

Built around GK01 riser and adapter card pairings, for both m.2 -> Oculink (for Handheld and minipc pairings) and Pci x8 -> SlimSAS (for SFF pairings) since the riser hole pattern is the same, the PCBs are interchangeable. This would be a universal eGpu enclosure.

The idea of SFFpc utilizing GK01 SlimSAS was the foundation of my "Osmium" case design. The riser pictured is OSmeta SlimSAS 8i.
This is a 2.8L APU build which in theory can link up to Gen4.0 x 8 lanes.

 

[SFFMunkee]

I have not had time to work on it unfortunately. I think that OneDock is a pretty good fit for when portability is required though.

For stationary eGpu installations, I have thought of a better solution using OSMeta GK01 which I have the hardware for but no time to finish the design. Basically would be four different enclosures of slightly sizes and gpu length, with flexible psu support for HDplex, Flex, SF and ATX enclosures. Then the user can optimize for card size.

Built around GK01 riser and adapter card pairings, for both m.2 -> Oculink (for Handheld and minipc pairings) and Pci x8 -> SlimSAS (for SFF pairings) since the riser hole pattern is the same, the PCBs are interchangeable. This would be a universal eGpu enclosure.

The idea of SFFpc utilizing GK01 SlimSAS was the foundation of my "Osmium" case design. The riser pictured is OSmeta SlimSAS 8i.
This is a 2.8L APU build which in theory can link up to Gen4.0 x 8 lanes.

That’s wicked, I like it! I kinda want to build one with an x8x4x4 bifurcation for moar SSD and eGPU. Not because it’s worthwhile but because it’s possible haha

 

[Alucard]

This is consistent with the trends I recorded on Gen 3.0 which showed only 3-4% losses at x4, and 0% losses at x8 comparatively to Gen 3.0 x16.

Can I have a link to your measurements?

 

[msystems]

My tests are from testing in Timespy with the b660i board and comparing m.2->Oculink performance with x16 performance baselines, where I had some results within margin of error. However there are a few tests done by other sites which are good starting points also:

General observations:
1) Lower tier (slower) cards benefit less from more pci bandwidth
2) The higher the resolution (the lower the FPS) the less impact of bandwidth

Scaling for 5700xt:

PCIe 4.0 and PCIe 3.0 - Difference between x8 and x16 with the fastest cards - Where does the bottleneck begin? | Page 7 | igor´sLAB

Especially after the tests of the B550 motherboards, the justified question about the problem with the split PCIe 4.0 ports naturally arises. PCIe 4.0 with 16 or 8x connection should not show any…

www.igorslab.de

With a weaker card (up to ~RTX 2080) there is potentially less than 1% loss between Gen 3.0 x8/x16

Scaling for a 3080:

NVIDIA GeForce RTX 3080 PCI-Express Scaling

NVIDIA Ampere finally brings PCI-Express 4.0 support to the high-end graphics market. The new interface promises twice the bandwidth of PCI-Express 3.0. We've setup an AMD Ryzen 3900XT system to test how various PCIe generations and lane widths affect gaming performance.

www.techpowerup.com

4.0x4 and 3.0x8 within ~4% of Gen 4.0x16 and within ~3% of Gen 3.0x16.

Scaling for a 4090:

GeForce RTX 4090 and Core i9-13900K PCIe scaling test shows 2% difference between x16 and x8 mode | igor´sLAB

PCI Express (PCIe) is the standard interface for connecting devices within a computer, and the fourth generation of this standard, PCIe Gen4, offers significantly faster speeds and much greater…

www.igorslab.de

Here we have 4.0x4 and 3.0x8 within ~6% of 4.0x16.