Calculating the key OC points on the 7800

[HeavyH20]

After some testing, the best points on the core clock of the nVidia 7800 series cards are whole multiples of the core frequency oscillation of 27 MHz. That calculates to the following root core clocks:

(17 x 27) - 40 = 419
(18 x 27) - 40 = 446
(19 x 27) - 40 = 473
(20 x 27) - 40 = 500
(21 x 27) - 40 = 527
(22 x 27) - 40 = 554
(23 x 27) - 40 = 581

And, here is how a single MHz below the whole number threshold affects 3Dmark05 scores:

445 8052
446 8222

472 8404
473 8508

499 8669
500 8753

Now, what about the other two cores. How are those controlled?

Well, when you adjust the root clock, you actually adjust all three clocks.

The root clock gets bumped by 40 MHz to represent the real geometry or vertex core clock. So, lets pick 450 core, which so many vendors use. 486 is the best vertex clock value (18*27), which translates to 446 root (486-40). So, we would actually have 490 using a 450 root clock, a 4 Mhz pad on the required frequency.

But, the shader and ROP cores start at 415 (really 418.5) in low power, and jump in whole frequency values (ie, 17 x 27 = 459) to a value that is less than 13.5 MHz over or equal to the root frequency. In this case the root frequency is 450 which is 18 MHz over 432 (16 x 27) and 9 MHz shy of 459 (17x 27), so 459 is the ROP/Shader clock.

So, a little confusing, but you do have some control over the other two clocks.

So, here is a list of primary - vertex based (red) and secondary - ROP and shader based (blue) root clock OC targets. The red will yield the larger gain, and you will see a second, albeit smaller gain on the blue clocks where the optimal memory frequency matches the root clock.

419
432 - the nVidia spec root clock
446 - so many vendors use 450
459 - why BFG uses 460
473
486 - why eVGA, XFX, and BFG use 490 on limited cards
500
513
527
540 - How I got 9539 on 3DMark05 with a lowly 3200+ and one 7800 GTX
554
567
581
594
608
621

[pduan87]

Nice work. Just like the GTX, it has the "stepping" type of o/c.

[alexio]

At 447 the score will be lower then at 446?

[Ackbar]

Looking at my clocks... I can't seem to actually reach 473+ unless I hit 475 on the clocks, then it hits 473+ on RivaTuner monitoring. Something to keep in mind. I think this is why Anandtech and others have seen a jump at "475" which is erroneously reported and is actually 473.

[Kunaak]

excellent work there.

I can confirm for a fact this worked for me.
great method for finding the absolute best OC Speeds.

I'll make it a sticky so others can use this for sometime

[HeavyH20]

At 447 the score will be lower then at 446?

No, you will actually gain a little, but is it pretty flat. You get marginal gains as you bump the clocks, but, when you hit the integer value of the 27 MHz oscillation, you get a significant step in performance. That is what I was trying to illustrate with the single MHz bump.

Looking at my clocks... I can't seem to actually reach 473+ unless I hit 475 on the clocks, then it hits 473+ on RivaTuner monitoring. Something to keep in mind. I think this is why Anandtech and others have seen a jump at "475" which is erroneously reported and is actually 473.

Yes, Coolbits is always been off a MHz or two. RivaTuner is more accurate. And, you can also use the newer version of ATI Tool for nVidia cards (0.25 beta 2 and up). That is also more accurate, and, you can get a temp monitor for your video temps in Motherboard monitor. It is kind of cool.

I can confirm for a fact this worked for me.
great method for finding the absolute best OC Speeds.

Thanks, I thought is was a very handy discovery. It finally explains the strange jumps people had been referring to. And, as an FYI, Rivatuner 15.7 will address the multiple core frequency detection. So, you get to see all three cores in action.

[toledo]

Asus Extreme is clocked at 486Mhz .....now i know where these number came from

[massman]

well I have to say, impressive.

How long did it take you to find this out

[HeavyH20]

After reading a number of reviews about OC plateaus on the 7800, I needed to know more. The first hint was the 40 MHz bump on the clock for 3D apps. The second was a peak at Rivatuner 15.7 monitoring results. Then, 150 runs in 3Dmark05 to give me more data. So, one long weekend.

[Mr. Tinker]

Anandtech found that the "plateaus" occur at different speeds on the GTX than on the GT.
http://www.anandtech.com/video/showdoc.aspx?i=2519&p=3

[wkmgt]

Above and Beyond The Call of Duty-----Thanks!!

[don_vercetti]

i got a big jump at 474 mhz over 472...so something seemed to be happening there.

And this is great work, thanks a lot. very good indeed.

[HeavyH20]

Anandtech found that the "plateaus" occur at different speeds on the GTX than on the GT.
http://www.anandtech.com/video/showdoc.aspx?i=2519&p=3

I am glad to see it help people out. I have e-mailed the Anandtech video review folks the findings as well. Have to figure out the GT anomally, however.


The Anandtech findings are interesting. I saw some strange things with the reported clocks on some GT RivaTuner reports. Assuming 27 Mhz is the same oscillation value for the GT. The true base clock of the GT would be 405, not the reported 400. So, when you overclock the card to 478, which is 78 MHz above the reported clock, you are possibly off by 5 Mhz, and actually at 473 (478 is 73 MHz above the REAL clock). Maybe the BIOS games the reported clock. That is why the OC GT's are at 425 and 445 versus 430 and 450 in the GTX world.

Here is a RivaTuner output:



This is for a BFG GT OC. The stock clock is 425, and the geometric delta clock is supposed to be 40 MHz. So, why does it appear to bump 47 MHz? The real clock of the GT may be 432, not 425. That could explain the jump to 472 from 425. A 47 MHz bump instead of the expected 40.

So, it is possible all GT reported clocks are under by 5 MHz less than what you see in CoolBits. I will verify all this info on a follow up.

[mike]

At 447 the score will be lower then at 446?

no - not unless the actual core clock remains the same

[PlatoonSgtElias]

What i dont seem to understand is how noone is getting the mem higher then me on air .... i got my memory to run at 1750MHz... core was 633 but i had a ovp on at 1.6v

[rick_fx]

So, here is a list of primary - vertex based (red) and secondary - ROP and shader based (blue) root clock OC targets. The red will yield the larger gain, and you will see a second, albeit smaller gain on the blue clocks.... 554

So that's why my card is blocked at 553Mhz

[Mr. Tinker]

I just want to know what the plateaus are on GT's so I know what to shoot for.

[dwwalker14335]

They are the same as the 7800 GTX, but off by 5 Mhz, it appears. I suspect it has to do with losing 1/6th of the pipes. But, instead of 473, it is now 478.

[Mr. Tinker]

They are the same as the 7800 GTX, but off by 5 Mhz, it appears. I suspect it has to do with losing 1/6th of the pipes. But, instead of 473, it is now 478.

so look at the gtx plateaus, and add 5 mhz. Werd. thx.

[IYP]

very nice work this will help me alot thank you

[einCe]

if i saw this sooner it would have sped up my gt purchase quite a bit lol. invaluable info.

[freecableguy]

To me it looks like the ROP and shader run in multipliers ONLY while the geo domain is floating at whole numbers plus the delta (and easy way to get more power via BIOS mods). As long as the geo value is not more than one half of the base frequency from the ROP/shader value (0.5 x 17 = 13.5Mhz). Therefore if you get shader/ROP corruption at the next step value but you know the core can handle more geometric strain just edit the BIOS to get a larger delta (+50 like Asus? + 60? +100?). It would be interesting to see the 3DMark05 before and after scores for this....

[saaya]

so the delta can be edited in the bios?
what about the base clockspeed of 27mhz?
(pretty much the synonym for a cpu multiplier)
can it be changed?
maybe with a hardware mod?

afaik all cards work like this but use a much lower base clockspeed than 27mhz, i wonder why nvidia went for this high base clockspeed

[cadaveca]

so the delta can be edited in the bios?
what about the base clockspeed of 27mhz?
(pretty much the synonym for a cpu multiplier)
can it be changed?
maybe with a hardware mod?

afaik all cards work like this but use a much lower base clockspeed than 27mhz, i wonder why nvidia went for this high base clockspeed

Look at the crystal on the board... ...same reason the x800xl doesn't clock worth a damn.

Yes you can edit the delta in bios.

[Ubermann]

Some companys is giving out new GT bios with no delta clocks.
Dont really know why.