After spending time today talking in depth to a couple of MSI's P45 engineers and considerably more to Tony Leach from OCZ Technology, who spends a lot of time QAing BIOSes for companies like Asus, DFI and MSI, it looks like overclocking is going to become an increasingly more complicated art with the release of Intel's P45 chipset, as it mirrors many of the tweaks the current high-end X48 chipset offers.
If you want to buy and overclock on an Intel P45 board, Leach believes that you must learn how to use:
* GTL Reference Voltages
* CPU VTT and its relation to GTLs
* Clock Skews
* CPU PLL Voltages
This is because we’ve reached such a stage with the front side bus that the frequencies are getting easily out of sync. You can’t just throw voltage at things any more – that will only get you so far.
It’s a case of spending a lot of time increasing the CPU VTT (not over 1.35V – you’ll kill the CPU) and CPU PLL (not over 1.7V, because again you'll kill your CPU) and tweaking the GTL Reference voltages for the CPU and North bridge to be around 61-63 percent of VTT for 45nm CPUs and 67 percent for 65nm.
This is particularly noteworthy with quad-core processors because, if you’re finding core two and three drop off under Prime95 load, it’s down to the fact that the two CPU dies are not identical and while core one and four can hit the FSB you’ve set, cores two and three are having trouble. Tweaking the GTL can sometimes give them better stability.
Leach even went as far as to say you’ve got to find points on the board and check the actual voltages with a multimeter, because we’re talking some extremely minute changes and if there are elements of vDrop from the board or vDroop when the CPU loads the BIOS can be inaccurate.
Next you HAVE to play with the clock skews – MSI will have these on its P45 boards, Asus has them on its current X38 and X48 boards, while both DFI and Gigabyte also have them on their X48 boards. Basically as the data has to jump from the front side bus domain to the memory domain this window becomes smaller and prone to more jitter, the faster you increase either the front side bus or memory frequencies and timings. By adjusting the skew you can realign these clock signals and suddenly stability should return again – a good board will have less jitter in its signal generation and finer skew adjustments than one that's not as good.
The kicker that this is a completely blind art – you’ve quite literally got to sit there for hours and tweak the nuts off the board trying combinations of GTL and Skew settings until you find something that works. To make matters worse as soon as you change the front side bus, memory timings, the CPU (no two CPUs are identical, even if you buy a “Q6600 G0”), the memory sticks (there are different tolerances between batches of the same product, never mind different products!), update the BIOS, or even if you’re using the same board as someone else there’s no guarantee that one set of settings will work on another board.
So there you have it, prepare to invest in some serious time if you’re upgrading, or wait and see what happens when everything gets integrated on the CPU with Nehalem because AMD doesn’t have this problem. However word is on the net that Intel will lock overclocking out of every Nehalem CPU apart from the most expensive Bloomfield options, basically screwing over all us value-enthusiasts. With this in mind, is there even any point in review sites “reviewing” overclocking any more, since it's unlikely that the end user will be able to replicate it because of equipment diversity?
