I have fixed the problem, just a quick recap for anyone running into the issue.
According to what i have read it has to do with the pcie address space that by default only ran to 32bit on this motherboard.
The pro duo cards have a pcie switch and the two gpu's it self so they consume some address space.
Also the board was crammed full of nvme and other stuff, so the third card could not be addressed.
You have to tweak the following settings in bios if you run into this:
Disable CSM
IOMMU on
Above 4g decoding <------ most important one to turn on.
After that it worked like a charm
How cool is it to see the following line in the software:
running the test data set on up to 32 CPU cores and on up to 6 GPUs
Hmmm, I hope you're right but i think GPU's are a lot harder to chop up into chiplets than CPU's.
Maybe tearing of the controllers and some cache of to a separate die could be an option, but separating shader cores to different chiplets would result in a crossfire like solution with dies, that nobody would be happy with.
GPU die bandwidth is so much higher than a cpu, and also shader cores are much more intertwined in processing than cpu's because of the inherent type of processing and sensitivity to latency.
It would only work if the os see's a gpu chiplet as a coherent GPU, and i don't think interconnect technology is that far that this can be done without great inefficiency and loss in performance in areas like gaming.
In compute maybe.
(probably why there are lots of whispers that blue team's gpu in performing way below expected efficiency)
Would love to see a 8 core zen CPU chiplet+ IO die, navi GPU , and HBM stack as L4 cache on the same chip 
In principle it would be more easy to make a gpu than cpu on a new process, the die size might be bigger on average gpu but the transistor density on a cpu is much higher. A gpu often has a lot less transistors per mm2, and because of a large die the heat is also less dense.
This is the reason that a gpu does fine with paste and little gain with liquid metal, because it already has a large surface.
Cpu's have a much more dense construction and heat spots, thats why the solder with cpu is by far the best way to go.
And if you cant go faster, go wider.
Also the nodes from foundry's are not directly comparable, those sizes are just marketing terms.
To be fair i had a Radeon VII, sold it within 6 months. Awesome compute card but very hot, loud and unstable for 24/7 usage.
I dont hold raja in high regards, probably a nice guy but i think hes not the guy you want in the leadership role.
I dont think team blue gpu will be anything good.
With all the XE gpu talk from him, i always think back to...... Poor volta
Lisa su made the right call investing the little money there was in CPU/ryzen development, and she keeps all in good order.
Now that the sky is the limit investment and knowledge will be returned to gpu's as well.
With the upcoming exascale computers there is also massive funding for the radeon instinct software ecosystem, this will eventually also trickle down to consumer stuff.
I started to use liquid metal as TIM in laptops, insanely efficient runs 20 degrees cooler in an instant.
Now days do everything with it, laptops, cpu, gpu. Just as easy as paste but always more efficient
