Diamond Plus SLI Mysteries starting to expose themselves.... LONG

Question

Arg...  to many shots in the dark with folks' SLI problems and the new SLI X16 getup .  Probably not the power supply, nor the video cards, not the memory, nor anything else obvious.  It may be those things, but after weeks of experimentation, there seem to be lots of other things to look at as well.   I've discovered several things about this mobo that didnt appear obvious in the beginning.
The obscure BIOS settings seem to matter in SLI mode, not in single card mode.  I think I'm starting to understand why.  Also, most of the drivers that are out there from Nvidia simply suck, and the right combination has to occur for the whole thing to remain stable, especially in SLI mode. WHQL has NOTHING to do with it, which I discovered after much pain and deinstalls and reinstalls.
The board comes set up from MSI in a conservative state, not prepared for SLI, or at least balanced SLI performance.   This can be witnessed in the Hypertransport settings in the Cell menu, where all NON CPU specific HT settings are 8x8 and 800Mhz.  This is clearly a setting for ONE video card, as it will create an I/O imbalance between the regular SLI ASIC (now the new southbridge) and the new SPP (system platform processor in Nvidia speak) ASIC.  Each of these 2 ASICs provide 20 lanes of PCI express each.   Each Vid card needs a complete 16 unencumbered lanes from its respective bridge chip to function properly in SLI mode, cuz if not, then one will simply not keep up and cause corruption.
I set this up and repeated the experiments several times and found the following.   It appears that for SLI to function properly, the HT settings between the Northbridge, Southbridge and CPU must be set to 1000Mhz and 16 16, in order to balance I/O thoughout the system.  This makes sense as it could starve 1 of the Vid Cards for data, the one on the Southbridge most likely (methinks that this is true because of how the new Northbrige appears to be directly wired to the CPU, while the southbridge is now servicing all other I/O).  This is so because the Northbridge has 1 set of 16 lane PCIE and the Southbrige has the other set of 16 lanes.  This is how the SLI X16 is enabled to provide 16 full lanes to each Vid card. 
If 1 of the PCIE slots is pumping I/O at a full 1Ghz, 16 bits in each direction, while the other is choked at 800Mhz and 8 bits in each direction, then there will be starvation on the second slot and the card that is in it when it comes to rendering large scale images at high speed.  Makes total sense.  In the networking world, its referred to as under and oversubscription. 
However, in this case, it is unobvious cuz the MSI docs (asus, abit, indeed Nvidia too) all suck and don't tell us what to do or how to think about this stuff. 
As far as drivers go, that is a mystery train wreck as well. 
If you are basically familiar with NVidia drivers, then you know that they create all the building blocks, and then their OEMs take those building blocks and package them together to suit their needs.   The binaries of the actual drivers don't change, but definitions, information and config files change, as does packaging.  Thats why regardless of who provides the complete driver set, the Nvidia launcher starts at some point during the installation.   
What Nvidia releases on their website will most likely work with all Nvidia products, but not necessarily well.   The X16 mobo chipset has its own release, and I've noticed that some of the folks here have used the regular SLI driver set.  That may not work that well, or at all, because while the southbridge is identical to the regular SLI single chip, the northbridge is brand new, and somewhere in the driver, messaging between these two has to be arbitrated.   So that's the first issue.   
MSI ships mobo driver version 6.82 with the Diamond Plus, but Nvidia is currently presenting 6.85 on their website.  What's the difference?  It seems to be just the ethernet driver if you compare the packing lists (the SMBus looks different too, but I think its a typo as MSI's appears "newer" than the Nvidia one, which is simply not likely as its a silicon device driver that would come from Nvidia )
So for all intents and purposes, the two releases are the same if you do not use the Nvidia Gigabit Ethernet adapter (which I do not, I only use the Marvell, much faster and lower CPU consumption).  If you use the Nvidia GigE adapter, then there may be some kind of an upgrade here. 
Now on to the Graphics drivers.   Nvidia lists 81.98 as their current WHQL version.  This is a BFD version    WHQL means nothing really, as its a self certification in parts, and Microsoft cannot test all the components anyway, as there are zillions of driver types for windows.    There are currently about 30, (yes, you read that right THIRTY) known and CURRENT graphics driver sets for NVidia GeForce graphics ASICs (chips that is).  Ever hear of 82.65, 83.62, 83.40....  how about the hackers current fave 82.12.   Yup there are plenty of em.  Mostly OEM releases for laptops and stuff, but many of them have actual bug fixes in them, albeit you have to read through copius release notes to figure that stuff out.  (wow reading, what a concept)
I had bad luck with 81.98, but 82.12 worked better, although SLI performance became half of 1 vid card (not good, suppose to go the other way    )   But 83.90 is working well so far, even though its a laptop release apparently.   
So the short of it is, that we are on our own with this SLI scheme as we were before.  There is precious little information about X16, and most of the original SLI info is either dated, or marketing speak with little or no substance to the zillions of frustrated gamers who cannot get their stuff to work well. 
Hope this info is somewhat helpful, as I pieced it together from lots of disjointed data here and there, including technical briefs, press clippings, various articles and some websites.   I'd be interested if folks either agree or disagree with some of these assertions, or if you've found similar or contradicting info, as I want my system to be stable so I CAN KILL STROGG and the like 

Answer 1

Quote from: Bas on 19-March-06, 17:56:24
Many have tried, most failed.....just because those dual-12V-PSU's stink....
Try the OCZ Powermax 520W, yes that one, it has all the connectors en a nice SINGLE 33A @ 12V....that one will do the trick just fine.
Read this stuff, it's very related http://www.theinquirer.net/?article=29940
It's not the board, not the videocards, not the drivers....it's your PSU!, those multiple 12V rail ones plain don't have what it takes to handle such systems, that's the problem.
Minimum PSU Requirements from the evga website:
1.  7900gtx single - 450w psu with 22a on the 12v rail
2.  7900gtx sli - 550w psu with 34a on the 12v rail
3.  7900gt single - 400w psu with 20a on the 12v rail
4.  7900gt sli - 500w psu with 28a on the 12v rail
5.  7600gt single - 350w psu with 18a on the 12v rail
6.  7600gt sli - 450w psu with 24a on the 12v rail
Plus, the new nvidia cards require less power than the previous 7800 and 6800 series video cards.  Also if you decide on an ATI card(s), they require more power than these 7900 series video cards too.  There you have it in black and white directly from a video card manufacturer.
http://www.evga.com/products/moreinfo.asp?pn=512-P2-N570-AX
Regarding PSUs - I have now the OCZ 520w and Tagan 480w - both are solid and heavy and the shielding looks to be made of the same material:
OCZ PS 520w - 33a on the 12v rail, adjustable 3.3v, 5v, 12v rails, blue LED bling factor - 2 shielded 6-pin vga power cables and 2 4-pin shielded hdd cables, 1 sata power cable (2 connectors) - 23dba
Tagan 480w U22 - 30a on the 12v rail, option to split into 2 12v rails, no LED bling - 2 shielded 6-pin vga power cables and 2 sata power cable (4 connectors) - 22dba
Tagan 580w U22 - 36a on the 12v rail, same features at 480w U22 - 24dba
My recommendation, if you need the bling and power, the ocz ps 520w has it.  If you need more power, the tagan 580w U22 has it, if you need the least expensive psu the tagan 480w U22 is the way to go ($91) except it is not recommended for a SLI 7900gtx setup.