Beware 3DMark 2006 . . .

There’s an article floating around that I’m afraid you’re going to have to take with a big block of salt.

It purports to provide 3DMark 2006 scores for a K10 quad-core processor running at 2.5 and 3.0GHz.

The purported scores, K10 vs. K10, were:

2.5GHz: 23,768
3.0GHz: 30,031

Since these are higher scores than the current world record of about 27,000 achieved with a grossly overclocked Intel quad-core, this looks impressive.

There’s a few problems, though:

First, the 3DMark score went up 26% just by increasing the processor speed 20%. That’s a scaling of 130%. That’s pretty good, matter of fact, it’s too good to be true based solely on changing the speed of the same type CPU.

Second, this is especially true for 3DMark 2006, which is primarily a video card, not a CPU benchmark. As you can see here, the percentage increase in 3DMark scores based on increases in CPU frequency isn’t 130%, it’s less than 30%. For instance, for the quad-cores listed, a frequency increase of 25% yields just a 6% increase in the score. That’s a scaling percentage of 24%.

Therefore, an 20% increase in CPU speed alone ought to get you a 3DMark score about 5% better, not over 25%.

So where’s the rest? The article mentions that the video cards were overclocked. It doesn’t say if they were overclocked at 2.5GHz, but somehow I think not, so overclocking the video card would account for another chunk of that 26%, probably somewhat more than that of the CPU speed.

That still leaves somewhere around half the improvement unaccounted for. Where might that be?

It’s possible that “optimized” drivers were used to improve the score, but there’s a much simpler and probably more likely explanation:

3DMark 2006 can be run a number of ways, different resolutions and different video processing modes. Reduce the resolution and video processing and, as this article illustrates, the scores go up. The differences aren’t night and day, but the score can be manipulated enough to account for the rest of the reported increase.

I don’t people don’t like gnatty details, but unless you know the exact specifications under which 3DMark 2006 (and many other benchmarks) were run, and compare those numbers to those run under the same conditions, apples to apples, comparisons are likely to be useless.

Ed