• Welcome to Overclockers Forums! Join us to reply in threads, receive reduced ads, and to customize your site experience!

8800 Ultra Water Cooled Modded Test Results

Overclockers is supported by our readers. When you click a link to make a purchase, we may earn a commission. Learn More.

ViperJohn

Senior Radeon Mod God
Joined
Jul 5, 2004
Location
Santa Ana, Ca
Well since I put up test result for a 2900XT I though it only fair to put up the results
from an 8800 Ultra. Again my test rig is stuffing potatoes in the cards exhaust pipes lol. The
CPU is a Core2Duo and not a Core2Quad



Header.gif

3DM01-3DM03.gif

3DM05-3DM06.gif

Aquamark.gif


Viper
 
Do you happen to have any other 8800 benches lying around to compare? While impressive, it doesn't seem to be worth the price.
 
Hazaro said:
Do you happen to have any other 8800 benches lying around to compare? While impressive, it doesn't seem to be worth the price.

Not with the later and faster drivers. I put this up so people can see how the 2900XT 1Gig card results I put up
compares to it. While the Ultra is definately faster it certainly isn't $300+ faster lol!!!

With the 97.02 8800 drivers (that is the only drivers I have tested a modded 8800GTX with in the same identical
test rig) the 2900XT was equal to or faster (a lot faster in 3DM06) than a good modded and 245W TEC cooled
8800GTX in all tests except Aquamark and that wasn't a large difference either.

Viper
 
Nice temps and clocks VJ.

But is the 1024Mb GDDR4 OEM 2900XT really $300 cheaper than the Ultra, I would have thought that with a Gig of DDR4 ram it would hve been more expensive.

I know the standard 2900XT is good value as Ive been very impressed with DM20 results
 
Deanzo said:
Nice temps and clocks VJ.

But is the 1024Mb GDDR4 OEM 2900XT really $300 cheaper than the Ultra, I would have thought
that with a Gig of DDR4 ram it would hve been more expensive.

I know the standard 2900XT is good value as Ive been very impressed with DM20 results

Thanks. The VF-IV Black Diablo's are really getting the job done!

ATI is using the slowest speed grade (.9ns 1100Mhz) GDR4 that Samsung makes so it
is a cheap chip relatively speaking and probably costs less than the currently listed, top
of the line, 1.0ns, 1000mhz GDR3 chip.

The chips that have to be god awful expensive are the .8ns, 1200Mhz GDR3's on the ultra.
Samsung doesn't even show they are made so the speed bin yield on those is very low
and that means very expensive per chip.

I can't go into anymore detail about the pricing ATM.

Viper
 
Last edited:
ViperJohn said:
Not with the later and faster drivers. I put this up so people can see how the 2900XT 1Gig card results I put up
compares to it. While the Ultra is definately faster it certainly isn't $300+ faster lol!!!

With the 97.02 8800 drivers (that is the only drivers I have tested a modded 8800GTX with in the same identical
test rig) the 2900XT was equal to or faster (a lot faster in 3DM06) than a good modded and 245W TEC cooled
8800GTX in all tests except Aquamark and that wasn't a large difference either.

Viper

Get back to work..!!..:) :)
 
John,
Any chance at getting chilled water or a TEC through that block? If so, I may have you build me a Viper Fang 8800 Ultra.

Dom
 
dominick32 said:
John,
Any chance at getting chilled water or a TEC through that block? If so, I may have you build me a Viper Fang 8800 Ultra.

Dom

Yes a ViperVenom-V 245w TEC cooler. Pictures 8 through 11 at my picture site. Shoot me an
email if you want to talk further about it.

Viper
 
Last edited:
Hey VJ,
Whats your take on this, I recall reading it some months ago but as I didnt have a 8800 card at the time, I just moved on with my day :) .

In your testing do you think this holds true or is it Rivatuner reading the frequencies wrong ?

Ive been playing around with my memory clocks and going by Rivatuner my card clocks something like this.

(I only tested from 1137 to 1213 and all so note the MHZ may be off by 1 or 5)

1137 thru 1150 = 1134
1151 thru 1160 = 1152
1161 thru 1183 = 1161
1184 thru 1208 = 1188
1209 thru 1213+ =1215

What Im testing at this time is 730/1213 which going by this is 729/1215 actual


http://service.futuremark.com/compare?2k1=9211725

Any how I'd like your take on this if I can.
 
Last edited:
Oc'ing an 8800 definitely goes up in steps like that, for all 3 clock domains. Unless someone has a way to read the frequesncies off the board maybe we'll never know 100% but I remember reading this a while ago then experiencing it. It can be roughly verified by running benchmarks, one with a consistent score is best.
 
Deanzo said:
Hey VJ,
Whats your take on this, I recall reading it some months ago but as I didnt have a 8800 card at the time, I just moved on with my day :) .

In your testing do you think this holds true or is it Rivatuner reading the frequencies wrong ?

Ive been playing around with my memory clocks and going by Rivatuner my card clocks something like this.

(I only tested from 1137 to 1213 and all so note the MHZ may be off by 1 or 5)

1137 thru 1150 = 1134
1151 thru 1160 = 1152
1161 thru 1183 = 1161
1184 thru 1208 = 1188
1209 thru 1213+ =1215

What Im testing at this time is 730/1213 which going by this is 729/1215 actual


http://service.futuremark.com/compare?2k1=9211725

Any how I'd like your take on this if I can.

RivaTuner isn't reporting anything wrong. It's just reporting the number held in the registers via the drivers which is the set number and may not be the true clock. The clock generators stepping resolution determines that and the actual clock setting where the step will occur.

In the case of the 8800 Ultra's the true memory clock will always be an even multiple of 27Mhz which is the base crystal controlled frequency of the clock generator. If you set a memory clock in between an even multiple of 27 the clock generator will round it up or down to an even multiple of 27 for the true clock depending on where you are at in the range in between even multiples.

Now just to keep things interesting the true memory clock on the 8800GTX cards will step on some 1/3 and 2/3's multiple in between the even multiples of 27Mhz

The 8800's ROP core clock steps on multiples of 27 as well (and the Shader clock on multiples of 54) but there are some places in the settable core ROP clock range where you can get a true ROP clock that is 2/3 multiple in between even multiples of 27. Unfortunately the core clock generator isn't consistant about doing that. As an example you get true core ROP clocks of 702 (27 x 26), 720 (27 x 26 2/3), 729 (27 x 27) and 756 (27 x 28) but you can not get a 747 (27 x 27 2/3) true ROP clock at this point in the range. It is the same thing down at lower core ROP clocks. At some points you can get a true core ROP clock at a 2/3 multiple and at others you can't. With careful use of the ROP clock setting you can almost always get two different true core Shader clocks at the same core ROP clock though. RivaTuner's clock graph is a god send here. If you run ATItools Scan for Artifacts you can bring up RivaTuner then use it to change the cores 3D ROP clock in 9Mhz increments (Even, 1/3 and 2/3's multiples of 27) and then see the true core ROP and Shader clocks immediately in the graph. For the 8800GTX cards do the same with the memory as the GTX's will step the true memory clock at some 1/3 and 2/3's multiples. Just write down what set clock produces what true core ROP and Shader clocks (and the memory on the GTX cards) and you can save yourself massive amounts of OC trial and error time!

This contrasts to later ATI cards where the memory always steps on multiples of 9 Mhz and their clock generators always round down if you make a setting in between multiples of 9 on the memory. ATI cards always step the core on multiples of 6.75Mhz and again always round down if you are not dead on.

Viper
 
Last edited:
MadMan007 said:
Oc'ing an 8800 definitely goes up in steps like that, for all 3 clock domains. Unless someone has a way to read the frequesncies off the board maybe we'll never know 100% but I remember reading this a while ago then experiencing it. It can be roughly verified by running benchmarks, one with a consistent score is best.

You are actually talking something different from Deanzo's question. An 8800 has three memory domains with a domain physically being nothing more than 4 of the 12 memory chips in a domain group. Each domain has it's own register that holds the set clock frequency for that domain which RivaTuner can be setup to read. The belief was the clock generator actually worked that way and used the discrete values held in the three memory clock registers to independently set the true clock frequency of each domain. You always have to think about the set clock value held in a register and the true clock the clock generator sets as two different things as they can be VERY different.

The first drivers out for the 8800's (97.02's I think) set all three domain registers to the same value when you ran the memory clock up. The later drivers didn't (I haven't checked using the 150 series drivers). When you ran the memory clock up with the later drivers only the register for memory domain 0 (zero) would show the new higher set memory clock number while domain 1 and 2 stayed at the boot up 400Mhz clock.

Since the card benchmarks scores stayed relatively the same between the 97.02's, which set all three memory domain registers to the same (higher) clock value when you ran the memory clock up, and the next driver release (97.28 I think) which only changed the new set memory clock in the register for memory domain 0 it is pretty safe to assume the memory clock generator only used the clock value held in the domain 0 register for all three memory domains. If it didn't the scores would have dropped with 2/3's of the memory running grossly under clocked.

I always had a very hard time believing the difference in the set clocks in the domain registers was meaningful and they were a moot point. I doubt NV would have left it that way in the many driver releases that came after the 97.02's if it hurt the cards performance as it was easily correctable. IMO it was a case of RivaTuner being able to accurately display data held in a register that was not actually being used by the clock generator hardware and NV chose to not fix something what wasn't actually broken. Sometimes to much data is not a good thing lol.

Viper
 
Last edited:
Deanzo said:
Thanks mate

Very helpful :)

Yeah NV's true core clock stepping has been a bit of a head scratcher evern since the
7800-256 cards came on the scene lol.

Viper
 
Last edited:
Back