3DMark03 and HD 4870 X2 Performance Scaling Investigated

@ 2008/09/12
It's very hard to call certain scores bugged and others not just because less people are able to repeat the good numbers. As the cause of the low or high bugged scores is not known, I tried to find out which of the two is bugged.

As we all know that 3DMark03 is all about the GPU power, I ran through the HWBot benchmark data and filtered out all results since the Ati 2900XT of the high-end cards. Note that some data, such as of the CFX (which indicated 3x CF) may not fit the graphs because of the lack of results).

First of all, I wanted to know how the different cards, going from 2900XT to 4870X2, scaled over time in different benchmarks. I left out 3Dmark01 and Aquamark because those two are too cpu biased.



In comparison to single 3870:

After that, it was time to have a look at 3DMark03 in detail, especially to see how the 3DMark subtests scale.







Now, the problem with these graphs is that they don't really show the exact performance scalingThe only thing these show that the newer you card, the higher your score and the more cards, the higher your score.

That's why I re-used the data and analysed the way they behave in comparison to a single version of the series.



Now the problem with this graph is that it doesn't show how the cards scale in terms of multiple cores, but in terms of multiple cards. That's why I drew another graph indicating the scalability over cores.



As you can see, there's something seriously wrong with the X2 CF scaling: GT2 and GT3 are way too low! Why?

The first graphs in this post show the scaling how it should be, because I used all the data from HWBot that was available. The performance scaling graphs are in other words based on thousands and thousands of results. The subtest detail graphs are only the top5 of each category. Why? Because those are the results that are questioned. Now, in the X2 CF category only the score of Spl is bugged high, which means that the four others are bugged low. The incorrect subtest scaling is therefor visible, unlike what we see in the single X2 category, in which most scores are in fact not bugged. In other words: we should indeed be hitting 180k with X2 CF, not 160k.



This graph shows the poor GT2 and GT3 scaling with X2 CF. This is the data of the top5 in hwbot.

[b]Please know that scores made with the blackscreen bug are indeed bugged and shouldn't be made public!

No comments available.