I think 40xx series GPU’s are more demanding regarding PCIe. I saw something about bar issues in one post (for now the only devices that were affected on Orion were Nvidia 40xx or newer). Up to 30xx series seems to work even with Gen 4 as demonstrated by @ShivanSpS’s 3060 or my 3090.
Problems with PCIe Gen4 on the x8 slot!
Sure, I just wish it was possible to at least figure out what exactly the 40 series cards want that the Orion isn’t capable of. Especially when these cards work just fine even on old motherboards with PCIe Gen 2. Sucks that everything on this board is so buggy and opaque : (
1 Like
Keep in mind this is the first consumer ARM board where you can even plug in a dgpu and use it at all. We are just on the 2nd public bios, and we are not even sure we were suppoused to have access to it.
1 Like
I promised myself to be reader only. But level of bull shit went too high ;D
this is the first consumer ARM board where you can even plug in a dgpu and use it at all.
The first one was in 2014. And was named Mustang. Made by Applied Micro company.
We used it to make several kernel subsystems work on AArch64. PCIe, USB etc. I used graphic cards from several chip vendors in it. Matrox, ATI/AMD, NVidia.
Most of Linux distributions used them to build packages.
And there were several other systems since then with one or more PCIe slots to allow use of graphic cards and other extensions. Consumer systems. In wide range of prices.
3 Likes
You have knowledge of most of the Arm boards so you so know majority of the devices. While all the new comers have no idea about the times companies have tried to make consumer grade boards based on RISC architecture.
I am excited to see Arm or Risc-V succeed but not at the expense of open source developers getting burned out by vendors 
I am on read only mode since more than 6 months now. Happy to see CIX doing a good job so far with UEFI, waiting for them to convince upstream Linux maintainer to make acpi for Arm 
As you already know dtb is not the way forward.
Lets hope for the best. RISC to consumer grade in how many years? Idk
2 Likes
Thats not a consumer grade board. specially at $5000… At that point we could argue even Altera did it.
But if you want to go that road, we could say some Risc-V boards did it first, Visionfive and the K1 has a good enoght pcie implementation to run gpus, at least some of them.
Even the RPI5 may qualify, maybe.
2 Likes
Now that depends on whats you’re definition of consumer grade versus what is the English definition of consumer grade.
Anyways good luck.
Ok bye
Keep in mind this is the first consumer ARM board where you can even plug in a dgpu and use it at all. We are just on the 2nd public bios, and we are not even sure we were suppoused to have access to it.
Wow, excellent insight, thank you so much for sharing. What pisses me off isn’t even that the fact it’s hilariously wrong, so much as that even if it were correct your comment would still be just as unhelpful. Plus you’ve brought the bickering to this thread. Great job.
Now on a more productive note, does any of you with experience from other boards know anything about debugging PCIe bringup? I have zero experience with modifying EDKII or flashing this board’s firmware with customized bios, but willing to get my hands dirty. Unfortunate that Radxa doesn’t seem to be very interested in helping users with PCIe compatibility, but would be welcome to see that change…
1 Like
You are certanly not helping by calling me wrong and giving no reason as to why im wrong. I already able to do more with the O6 with the 2nd bios that showed up than in 4 years with the useless RK3588. and in 10 years of $10 to $200 SBCs.
Yes there are problems and i tried to help you, i dont have the tecnical knowledge to go deeper.
If you all rich people that can buy $5000 server boards and call it a consumer board want to call me wrong go right ahead.
1 Like
Mustang came to the market with price of 825 USD.
Which at that time was good price for what it gave.
@Salvador my comment wasn’t nice and I’m sorry for acting like an asshole, I was upset to log on and saw my troubleshooting thread turned into pointless off topic semantic arguments. My point still stands but I shouldn’t have acted like that. Sorry.
You are certanly not helping by calling me wrong and giving no reason as to why im wrong. I already able to do more with the O6 with the 2nd bios that showed up than in 4 years with the useless RK3588. and in 10 years of $10 to $200 SBCs.
@ShivanSpS Im confused, where did I call you wrong. Yeah this board is way more capable than cheap credit-card SBCs, I mean that’s the whole reason I bought it in the first place.
If you all rich people that can buy $5000 server boards and call it a consumer board want to call me wrong go right ahead.
You can pretty easily purchase boards like LX2 or put together used ampere for less than 1/5 that. But anyways this is the pointless argument I was trying to get away from, stop it already.
Yes there are problems and i tried to help you, i dont have the tecnical knowledge to go deeper.
I appreciate trying to help, even if I don’t understand how bridging the sense pin would do anything, especially when its an x8 gpu in an x8 slot anyways and the card is at least discovered in the devicetree image, and the 3.0x16 cards work at x8 perfectly even on old bios. But I’m down to try it if you think it would change something. I mean it’s the only suggestion I’ve gotten here…
Again I’d appreciate anyone suggesting how to determine what the 40 series are doing differently than my working ones.
There is a slight possibility that the BIOS update to come might solve this issue, hopefully, considering you have signaled this issue for a while. The BIOS that enabled the 30XX generation seems to be a bit pre-release, so who knows.
1 Like
One can hope. Looking forward to see the pcie power rail issue finally fixed too.
On the 9.0.0 SystemReady firmware, with an Nvidia 3080 Ti, I am only seeing PCIe Gen 1 x8:
LnkCap: Port #0, Speed 16GT/s, Width x16, ASPM L0s L1, Exit Latency L0s <512ns, L1 <4us
ClockPM+ Surprise- LLActRep- BwNot- ASPMOptComp+
LnkCtl: ASPM Disabled; RCB 64 bytes, LnkDisable- CommClk+
ExtSynch- ClockPM+ AutWidDis- BWInt- AutBWInt-
LnkSta: Speed 2.5GT/s (downgraded), Width x8 (downgraded)
TrErr- Train- SlotClk+ DLActive- BWMgmt- ABWMgmt-
See https://github.com/geerlingguy/sbc-reviews/issues/62#issuecomment-2855177172 for more
@geerlingguy try running lspci at load. It’s some power saving feature that downgrades the link but in my case it also upgrades the link when needed
1 Like
I was doing that; ASPM was disabled (which should mean the port should run at full speed, IIRC), and I also checked lspci under load, and monitored in real time with nvtop. It was always showing at 2.5 GT/s.
Could also be a quirk in the Nvidia driver… or with some userland applications (I didn’t re-check today while I was testing some LLMs, though they don’t need tons of PCIe bandwidth).
nvbandwidth is not usefull here to check the actual bandwidth under load?
Do you have any other x8 machine to test it on ? It’s possible that the card automatically downgrades to Gen1 when detecting it’s not in x16 mode. I would find this surprising for such a high-end board, but not totally unlikely. In the worst case if you’re feeling lucky, you can put some paper on the connector at the position that starts after x8 and insert it in an x16 slot. I admit that it’s not my preferred approach though!
2 Likes
@meco What did you mean with that? Did you see downgrade of PCI version or of speed due to power savings?
Does that mean ASPM is supported? If so it should be written up in the ASPM support thread. Ref O6 / Cix CD8180 supports PCI sleep states (Active State Power Management)?
Yes PCIe Version as well as link speed get scaled by need.
1 Like