I must agree too Z170 chipset is having problem with 3+ cards on most boards, its prolly bios related, but who ever is doing testing first disable all non needed IO and internal devices as:
- sound
- ports
- any add in hw as firewire, wifi, etc
- set PCI-e to x1 mode if possible
And only then test it, also better use win10 because it should handle 8 GPU with no moded drivers
Ill give it a try, however i just cannot wrap my head around the fact that the z170 chipset has more pciE lanes compared to the h80 or h97, in every single connectivity feature the z170 is superior to the h97, and somehow the former performs much worse than the later.
I guess that it is a hardware issue, but mostly a uefi issue, so maybe a custom rom or manufaturer update will solve this, however i wasnt able to fin any information of anyone trying to develop a custom rom for that purpose. And the bios updates do nothing regarding the 6 gpu compatibility issue.
This article alone won't help you fully uncover the whole problem (it's not even directly aimed at this exact issue of miners, just a general analysis):
http://www.tomshardware.com/news/intel-100-series-hsio-chipset,30210.htmlI think the author was on the right track in understanding the SkyLake platform's possible connectivity issues after piecing some important things together but he left the matter alone before a real final conclusion was formed.
I think the real problem is some undisclosed further limitation about how these HSIO lanes can be allocated to PCI-E or other devices:
As you can see, some HSIO lanes are fully or loosely pre-allocated by Intel. Some lanes can be configured as:
- USB only
- PCI-E only
- either PCI-E or USB
- either PCI-E or SATA
- Now look at that Gbe !!!
The weirdest thing on this graph for me is the limitation regarding the (built-in?) Gbe network controller's HSIO lane allocation. I would think it should either act like any other PCI-E device (freely allocated at any open PCI-E capable lane) or have it's unique fix preset like the USB or SATA controllers do (like USB#1 can only go to Lane#1 and SATA#7 can only go to Lane#20, etc). What kind of logic might be hiding behind that?
---> Take note about how you must pick 4 neighboring lanes starting at N*4+1 (like 5,6,7,8 but not 6,7,8,9) for PCI-E 4x devices.
Another strange thing here is how you can theoretically have as much as 5 separate 4x devices (if you are willing to sacrifice all SATA and Gbe but otherwise it's still 4 with 1 Gbe and 1+ SATA), yet only 3 of these kind of 4x devices are allowed to be used simultaneously.
At some point I though I figured this out: I can operate 3 PCI-E cards because I had 5 PCI-E 4x blocks, 1 allocated for SATA and 1 for Gbe (and/or some SATA at the same time), thus leaving me with 3 free PCI-E 4x blocks, all of which are capable of operating a single PCI-E device regardless if that is 1x, 2x or 4x
But... I also had an integrated USB3.1 and a sound controller which are both PCI-E devices! So, this 4x block = 1 device can't be true. Well, unless I was actually operating 2 cards on the PCH and the third one was in the 16x slot and it's wired directly to the CPU (not micro-managed by the PCH - but I think it is because the Z170 motherboards can be configured in 8+8 modes whereas H170 boards are limited to a single 16x slot but this could be an artificial limitation imposed at firmware leve but this would also explain why Z170 seems to handle 4 cards and H170 only 3 cards: it's 2 vs. 1 cards on the "direct" CPU 16 lines and always only 2 on the PCH).
And now, if you thought you almost figured this out as much as it's possible..., look at how SATA#0 and #1 can walk around the table! What does that supposed mean? At this point I am beginning to question if this table exists to shows the fixed limitations (more than the USB ones, that is) or it's merely an example which illustrates several possible (and probably recommended) alternatives at a single graph. So, my weak speculation is destroyed here.
------------------
If you are into some experimentation (either by virtue or out of necessity), it would be interesting to hear about any of these C232 or C236 workstation boards (yon don't need Xeon CPU and ECC memory, just plug in your cheap Pentium and regular non-ECC sticks if you wish):
http://asrock.com/mb/Intel/E3V5%20WS/index.us.asp (5 slots for moderate price)
https://www.msi.com/Motherboard/C236A-WORKSTATION.html#hero-overview (6 slots, expensive)
http://www.gigabyte.com/products/product-page.aspx?pid=5806#ov (6 slots, expensive!)