Can someone explain faster speed = lower 3dmarks??

E

Evoic

Gawd
Joined
Apr 30, 2007
Messages
856
I ran two 7900GTX's against each other to compare speeds and scores and to see if they came up any different.

The first card (Card "A") idles around 42c and loads at around 68c.
The second card (Card "B") idles around 37c and loads at around 62c.

They are identical in every way.
Stock cooling for both.
CPU Speed exactly the same as well.

Card "A" clocks at 710Mhz core / 1765Mhz Memory - 12462 score in 3dmark05
vs
Card "B" clocks at 730Mhz core / 1850Mhz Memory - 12412 score in 3dmark05

:confused:

How / why is this?
Any explanation would be appreciated.
If that explanation were kept simple enough for a 3rd grader to understand - that would be appreciated even more.
;)

Thanks.
.
.
 
You are pushing your card too much. You are actually losing performance because your card cannot handle it. PLUS every card is different.
 
Card B has a BIOS with relaxed timings, allowing for higher overclocks but less performance. It's a 'trick' commonly used by the board partners to allow them to clock the card's higher and advertise them as having higher clocks.

Joe Consumer sees the card has higher clocks and thinks "Ooooh more faster dudez!" and spends his money on that one.
 
aznx said:
CPU bottleneck.
Click to expand...

CPU Bottleneck with the same processor at the same speed, with both cards tested 45 minutes apart?

CPU bottleneck on an E6420 with 4MB's of cache clocked at 3333Mhz?


lloose said:
You are pushing your card too much. You are actually losing performance because your card cannot handle it. PLUS every card is different.
Click to expand...

This sounds reasonable.
I'm going to retest at the same speeds thatI ran the first card at and see what happens.


Blue Falcon said:
Card B has a BIOS with relaxed timings, allowing for higher overclocks but less performance. It's a 'trick' commonly used by the board partners to allow them to clock the card's higher and advertise them as having higher clocks.

Joe Consumer sees the card has higher clocks and thinks "Ooooh more faster dudez!" and spends his money on that one.
Click to expand...

I'm not sure if I'm understanding you correctly.

These are no cards using the same reference design, from different mfg's......

These are the exact same cards from the exact same mfg. with the exact same model number.


For the record, I had closed almost all apps running in the background with the first run with card "A" and had a few things open during the test with card "B".
 
Second for pushing it too hard. It still works, but generates slight errors, thus slowing things down. CPUs will display the same behavior if they are overclocked too far.
 
No, the REAL reason is because the benchmark can be off by a few hundred points if you run it multiple times, so your difference is small, and because you had more apps running during "B"'s test, less resources were available for 3dmark.
 
Caffeinated said:
Second for pushing it too hard. It still works, but generates slight errors, thus slowing things down. CPUs will display the same behavior if they are overclocked too far.
Click to expand...

Well, then that is the strangest part for sure.....

Card "A" was tested with ATiTool and went nuts with the errors if I went even another 1-2 Mhz higher on memory or core.

Card "B: had absolutely no errors at the speeds posted above (faster than card "A") and I simply stopped trying for higher because I wanted to see how much of a difference there was between those speeds and the speeds I maxed at with card "A".

?????
 
Timings are set in BIOS, so you hit the nail on the head when you said they were from different manufacturers. Different manufacturers, different BIOSes on the card, same reference design. Card A goes crazy if you clock it higher due to tighter timings. Card B will keep going up without issue due to looser timings.

It's like RAM. You can run DDR2-800 rated stuff a lot faster if you set it to like, CAS7 rather than the CAS4 or 5 it's rated for out of the box. The manufacturer of Card B loosened timings to make it more OC friendly, and just like DDR2-1066 at 7-7-7-18 is actually slower than DDR2-800 at 4-4-4-12, the second card with looser timings but higher clocks is actually slower than the first card with tighter timings but lower clocks.

Blue Falcon got it right.
 
i can tell you from years of using 3dmark as "some sort of metric" for OC's on all parts of the rig that were capable...you can run it 5 times in a row, and never get the same score. Maybe you forgot to kill a process that you had killed in the previous boot. Maybe your page file is slightly more fragmented one time, over another, maybe your file system. It doesn't really matter what the cause. With my rig, I have about 200 points leeway in my score, at all times. You'd be better off using a real world game bench for minute clock comparisons like this. My guess is that the difference is less than 5fps between the two, anyways.

The others may be right about the timing, etc. It makes sense (they do it with everything else, right?) but it's the first I'd ever heard/thought about it on video cards. It may be a factor, but I'd never heard of it before.

If I am not mistaken, our very own beloved [H] editors use multiple runs for all FPS and score tests when doing gaming benchmarks. It is quite simply impossible to recreate the exact same score everytime in 3DMark, I don't even care what version. I know they are able to use pre-recorded timed demo runs + FRAPS to get real game FPS (when a timedemo command cannot be used), but even this is surely the average of multiple runs. I would love to see an [H] editor chime in here and set the record straight.

KYLE!!!? Where the hell are you, man!?
 
silent-circuit said:
Timings are set in BIOS, so you hit the nail on the head when you said they were from different manufacturers. Different manufacturers, different BIOSes on the card, same reference design. Card A goes crazy if you clock it higher due to tighter timings. Card B will keep going up without issue due to looser timings.
.......
Click to expand...

These cards are from the same Mfg.

sabregen - Thank you for that post. That makes me feel at least somewhat better about what may be going on.

.....and to test the theory, I'm going to reboot - crank up the speed, shut down some processes, and retest.

BRB.
 
Previously:

Card "A" clocks at 710Mhz core / 1765Mhz Memory - 12462 score in 3dmark05
vs
Card "B" clocks at 730Mhz core / 1850Mhz Memory - 12412 score in 3dmark05

----------------------------------------------------------------------------------

2nd attempt, card "B"

Card "B" clocks at 737Mhz core / 1850Mhz Memory - 10658 :)eek: - oops......) score in 3dmark05


3rd attempt - speeds scaled back to match original Card "A" speeds:

Card "B" clocks at 710Mhz core / 1765Mhz Memory - 12466 score in 3dmark05 :)


It appears I was pushing too hard. :p
Now I need to add some aftermarket cooling and find the sweet spot.

Thanks for the tips, posts, and guesses.

This was an informative thread for me - I'm still somewhat raw when it comes to GPU overclocking.
 
And you are SURE you have the same processes running?

Run the test 5 times in a row and see how much it changes.
 
agreed, multiple runs to get an average. I am starting to suspect Vista Ultimate for the variances in my personal bench runs (yeah...like I shouldn't have known that before, right?)
 
Everything at the lower speed is coming up nearly identical.

When I push the card, it starts to give lower scores.
Not sure if it's throttling (don't think so, max temps are under 70c, but it's only running stock cooling at the moment), if it's pushing it past it's limit of efficiency on the memory side, or if the speed of the card is bottle-necked by the CPU.

I don't think the CPU is the problem, actually.
The CPU scores are consistent on nearly every run, it's only a steep decline in the actual graphics score when pushed to hard.

Oddly, there's not one single instance of artifacting in the demo or using ATiTool to check.

:confused:

Electronics.

Crazy stuff.
 
sabregen said:
i can tell you from years of using 3dmark as "some sort of metric" for OC's on all parts of the rig that were capable...you can run it 5 times in a row, and never get the same score. Maybe you forgot to kill a process that you had killed in the previous boot. Maybe your page file is slightly more fragmented one time, over another, maybe your file system. It doesn't really matter what the cause. With my rig, I have about 200 points leeway in my score, at all times. You'd be better off using a real world game bench for minute clock comparisons like this. My guess is that the difference is less than 5fps between the two, anyways.

The others may be right about the timing, etc. It makes sense (they do it with everything else, right?) but it's the first I'd ever heard/thought about it on video cards. It may be a factor, but I'd never heard of it before.

If I am not mistaken, our very own beloved [H] editors use multiple runs for all FPS and score tests when doing gaming benchmarks. It is quite simply impossible to recreate the exact same score everytime in 3DMark, I don't even care what version. I know they are able to use pre-recorded timed demo runs + FRAPS to get real game FPS (when a timedemo command cannot be used), but even this is surely the average of multiple runs. I would love to see an [H] editor chime in here and set the record straight.

KYLE!!!? Where the hell are you, man!?
Click to expand...

Good points, too. There is going to be some sort of difference; standard deviation of 5% should certainly be taken into account, which if we are talking a score of, for instance, 10,000, that would be 500 points either way.
 
sabregen said:
i can tell you from years of using 3dmark as "some sort of metric" for OC's on all parts of the rig that were capable...you can run it 5 times in a row, and never get the same score. Maybe you forgot to kill a process that you had killed in the previous boot. Maybe your page file is slightly more fragmented one time, over another, maybe your file system. It doesn't really matter what the cause. With my rig, I have about 200 points leeway in my score, at all times. You'd be better off using a real world game bench for minute clock comparisons like this. My guess is that the difference is less than 5fps between the two, anyways.

The others may be right about the timing, etc. It makes sense (they do it with everything else, right?) but it's the first I'd ever heard/thought about it on video cards. It may be a factor, but I'd never heard of it before.

If I am not mistaken, our very own beloved [H] editors use multiple runs for all FPS and score tests when doing gaming benchmarks. It is quite simply impossible to recreate the exact same score everytime in 3DMark, I don't even care what version. I know they are able to use pre-recorded timed demo runs + FRAPS to get real game FPS (when a timedemo command cannot be used), but even this is surely the average of multiple runs. I would love to see an [H] editor chime in here and set the record straight.

KYLE!!!? Where the hell are you, man!?
Click to expand...

I have not read this entire thread, but yes, 3DMark is what it is....but it is not an exact scientific tool not matter how many others tools try to get you to think so. ;)
 
jmackay said:
No, the REAL reason is because the benchmark can be off by a few hundred points if you run it multiple times, so your difference is small, and because you had more apps running during "B"'s test, less resources were available for 3dmark.
Click to expand...

QFT. You'll rarely, if ever get identical results even on the same machine. Run these tests five times on each machine then take the average. I would call that your closest, best result.
 
No one is paying attention here.....LOL.

READ THE FIRST POST.

I was experiencing slightly slower scores on card B - even though the speeds were noticeably higher in the first run.

In the second run, with a few more Mhz on the core - the performance went WAY down.

I switched the speeds on card B to match the original speed of the very first run on card A that I was testing against and voila'......the score came out the same.

A quick recap.

Card A - 1st run - solid score
Card B - 1st run - faster speed / slightly lower results. :confused:
Card B - 2nd run - even faster speed / extremely reduced results. :eek:
Card B - 3rd run - speed matched card A above / score was nearly identical as well. :)

I learned:

  • Both cards score almost identically.
  • Card B suffered lowered performance because it was pushed too hard. ;)
 
Evoic said:
Card B suffered lowered performance because it was pushed too hard. ;)
Click to expand...

I said that in the 2nd reply to this thread AND the op even agreed with me.. At that point, it should have been /thread.. Kyle just added that 3dmark is a piece of shit for benchmarking (in nicer words).

You're right, somebody isn't paying attention.. :rolleyes:
 
...and perhaps that someone was me, I suppose. I know what you were going for, but my point is that anything within 5% is covered in the standard deviation that is considered to be acceptable variation in the score. You cannot simply take run A and compare it to run B, if you're truly after results. What you should do is make a batch of 3 or 5 runs at each clocks speed, average out the runs in the batch, and then compare them. Comparing one run against the next is a misleading metric, and not a good sample of the systems score at a given clock speed. You must do average versus average to gain a truer metric, plain and simple.

Now, that being said, if you're just doing constant runs of a benchamark, and in between runs, you're clocking a component up and up, until the point of failure, that's fine. However, that scenario (make a 3DMark run, push the clocks, run 3Dmark, push the clocks) should ONLY be used to find the maximum OC on a component before a hard lock/BSOD/no video error occurs. It's good for finding the MAX that you component is capable of. Once found, start working backwards until you can run multiple times (while staring intently at the screen, acnning for artifacts) without crash to desktop, or video driver crashes. Once you've acheived that sort of stability, it's time to start doing your bench runs in batches, and comparing each batch of multiple runs against eachother.

But I think I'm going to be flamed for still missing the point, and beating a dead horse here. All I'm saying is average out several runs at any given clock speed, and record the average when doing benchmarks.
 
This is why anyone with any common sense isn't running 3DMark as a judgement for differences in hardware/clocks. The benchmark itself can have a good few percent margin of error which at the scores you're looking at equates to a good few hundred points total. If you want to increase that accuracy then run the benchmark 5+ times for each card and average the scores (mean)

Otherwise grab something else more appropriate such as an actual game, and use a built in time demo or benchmark (like FEAR)
 
Frosteh said:
This is why anyone with any common sense isn't running 3DMark as a judgement for differences in hardware/clocks. The benchmark itself can have a good few percent margin of error which at the scores you're looking at equates to a good few hundred points total. If you want to increase that accuracy then run the benchmark 5+ times for each card and average the scores (mean)

Otherwise grab something else more appropriate such as an actual game, and use a built in time demo or benchmark (like FEAR)
Click to expand...

How many different ways do I need to say this?!?!?! :mad:

When I ran one card a little hard, the performance went down a bit.
When I ran that card too hard, the performance went down drastically.

ASIDE FROM THAT - THE PERFORMANCE WAS NEARLY IDENTICAL FOR BOTH CARDS.
 
You must log in or register to reply here.
Back
Top