9600GT core clock mystery seem to be revealed

Gentlemen, it looks like we can finally close 9600GT core clock mystery, which I’ve been brainstorming with my friends during the last 2 weeks. The mystery was not visible to public, however almost all reviewers noticed it during monitoring 9600GT clocks in realtime. Unfortunately, it was not exposed in majority of 9600 reviews, but TechPowerUp’s Zotac 9600 GT review is a good example of review which focused attention to it (please refer to “Wrong core clock?”):

http://www.techpowerup.com/reviews/Zotac/GeForce_9600_GT_Amp_Edition/24.html

Summarizing:

NVIDIA insisted that they use 25MHz crystal for 9600GT’s core clock generation, which sounded a bit suspicious for us because it is simply not logical to use two different crystals on the same PCB and furthermore we couldn’t locate 25MHz crystal on the PCB. NVIDIA also refused to comment our questions about lack of 25MHz crystal on the PCB.
However, yesterday I’ve found really interesting report in the forum, which pushed me in right direction and shed some light on the mystery. One of 9600GT owners (kudos to Gljukats @ overclockers.ru) reported that his 9600GT is really sensitive to PCI-E bus clock tweaking and even minor PCI-E clock increase increases 9600GT performance in 3D benchmarks (which are normally not limited by PCI-E bus bandwidth). This allows me to make the following conclusion: there is no fixed 25MHz reference clock and NVIDIA uses quarter of PCI-E bus clock (100MHz / 4 = 25MHz) as reference clock for core frequency generator. I’m almost 100% certain that my assumption is correct, however I still don’t have 9600GT installed and cannot verify it myself. Anyway, I hope that I’ll be able to do it really soon.
So 9600GT owners should keep in mind that GPU clock is linked with PCI-E clock on these cards. So for example by overclocking the bus from 100MHz to 120MHz you also overclock your 9600 GT GPU by 20%. And the most tricky thing: you won’t be able to see clock frequency changes either in NV driver or in RivaTuner or any other software, because NVIDIA always uses fixed 25MHz crystal clock in the calculations.

 

Don't think that they'll change it. Reading true GPU clock in this case means that the driver will need to read PCI-E clock directly. That's a personal codepath for each mainboard chipset and even many mainboard models, using external clock generators. Display driver is not a place to implement such things. Most likely they will always suppose that the clock is default and use fixed 25MHz for core clock calculation. I'll also do the same in RT's hardware monitor, I physically won't be able to support dozens on mainboard chipsets and models because RT is not oriented on low-level mainboards support.

No. 2.07 will be launched as soon as my 9600GT sample arrives. I'm already waiting for it during almost 2 weeks.

Yes.

 

The hypotesis has been confirmed by W1zzard's benchmarks:

http://www.techpowerup.com/reviews/NVIDIA/Shady_9600_GT/

 

I guess that they tried to hide it from the competitors, not from the end users. Just like it was with independent domains clocking a few years ago, which we've discovered after 7800GTX announce.

I'm pretty sure that will apply to both reference design and non-reference design 9600's. Cannot say anything about 9800 until seeing a sample.

I'm afraid that there is some misunderstanding. What makes you think so?

 

Sorry but, exactly where do you base that info on ?

I just checked on my 780i (eVGA Ref) and it's 100 MHz, if you check any review you'll notice that all of them show a 100 Mhz reference speed.