Product: ATI HD 4870 X2 QuadCF
Company: AMD
Authour: Alex 'AlexV' Voicu
Editor: Charles 'Lupine' Oliver
Date: December 24th, 2008
A Merry Introduction

December's usually a gift-loaded month, with red and white the two colors intrinsically tied with these ubiquitous gifts.  Already well into December, it becomes quite clear that we should be sending a red and white dressed gift your way - and so we shall!

Dave 'Wavey' Baumann holding his boomstick
Dave 'Wavey' Baumann holding his boomstick

A great gift must be both expected and surprising - not an easy mélange at all; luckily for us the right ingredients were within hand's reach! On one hand, the dual 4870X2 configuration that's going to sweat itself to death (hopefully not, they're cute cards) rendering pretty pictures is the expected part - after all we alluded to this review in the closing statements of the 4870X2 investigation quite a while ago. But wherever could a novelty be extracted from? After all, video card reviews are a rather typical affair ... or are they? Stick around and you'll see that even old dogs can be taught some new tricks.

Alongside the “meaty” portion outlined above, additional sprinklings of spices will be added in order to complete the intricate dish - namely, we'll be looking at the 8.12 Catalyst drivers, which are a milestone build, of sorts, with the obligatory GPU accelerated transcoding of the ... ”educational” movie collection included.

Now, be forewarned that this will be one of those articles that few but the author likes, because it will be long and will include rather detailed explanations that are seldom interesting to the vast majority ... we've tried to balance this flaw with liberal use of rather pretty graphs, but it remains to be seen if they'll prove sufficient.

Finding the right mix ...

Is damn hard, and not only when it comes to coffee or music! Even reviews are hard to get right, and we've yet to manage one that is actually as it should be. Delving into the hardware itself at this point in time would be something less than wise; after all, little has changed on that front since the 4870X2 reviews have hit the Interwebs and all details should be well known by now. Quad-Crossfire itself is a fairly old acquaintance, which means its ins and outs won't surprise many self respecting enthusiasts. The testing methods are also typical and commonly employed ... oh, wait, they're not! Actually, they're quite different from what we've been doing, or have seen done, and hopefully they'll end up near the superior edge of the [suck, rock] continuum. So what's this nonsense all about you ask?

The general practice, and one that we've followed up until now, is to provide average, minimum and maximum framerates. Whilst this is a good start, it has become quite apparent that those indicators are hardly sufficient for painting a complete picture: the average itself is quite sensitive to outliers (extreme values that are not necessarily representative for the run/dataset, but impact the average), and minimum and maximum framerates can well be a one time thing. Imagine you'd hit a minimum of 1 fps, but that happens exactly once in a 4 minute run, with all other frames being >60 - how would one quantify that particular case, based solely on AVG, MIN & MAX? As a consequence, we've taken number of steps to ensure that the picture painted via testing is far more representative.


The first, “sharpen-filter”, is composed of 3 extra statistical indicators:

  • Median: going by the book, this is the number separating the higher half of a data set from the lower half; what this means, in our case, is that at most ˝ of the recorded FPS values will be less than the median, and at most ˝ of them will be more than it (the way of calculating it, whilst simple, is beyond the scope of this article).

  • Mode: by definition, this is the value that occurs more frequently in a data set (be aware the mode is not unique, more values can have equal frequencies); in the present situation, it'll reflect the most frequently “experienced” FPS ... which is to say, where the frame-counter most know and love so much stayed the most often.

  • Standard Deviation: finally, one of the most commonly used statistics tools, which measures how widely spread the values in a data set are; you could regard it as a way of gaging the “quality” of the average, or, in other words, determining just how good the average itself actually is at representing the data set it's based on, whilst also being useful for indirectly gaging the amplitude of the FPS swings around the average (both higher than and lower than the average deltas are accounted for)

Of course, a picture is worth a thousand words ... so we'll include two and save two thousands:

Standard Deviation
Standard Deviation

What you see on the left is what happens when the average is affected by an outlier: the data set we used was {25,25,25,100}. The average, in this case, is equal to 43,75: hardly representative for the experience. However, the Median and the Mode are both equal to 25, which produces a far more accurate synthesis: we know where half of the data set lies, that at most half of the values will be less than this, and we also know that most of the time the user experienced 25 FPS.

Moving to the right, you see three data sets that share the same 25 FPS average, but the end-user experience between the three would be quite different indeed:

  • The Red Line: On one end of the spectrum, the ideal case is represented by the red line, with no variations. However, the likelihood of achieving that in practice is reduced, since there are framerate variations inherent to the rendering process, caused either by the differences in rendering load associated with moving through scenes, by resource uploading as one moves through the virtual world, or a number of other factors.

  • The Black Line: What is possible, and desirable, in practice is the situation presented by the black line, with values that are rather tightly packed and close to each other and the average.

  • The White Line: The white line serves the purpose of presenting a bad situation, where the average tells about zilch about the experience itself, since swings are so ample. Take home note? Low values for Standard Deviation are desirable.

Using the above three indicators should add more weight to the measurements we currently employ for gauging a GPU's performance. However, the exploration must go further, and to do so a histogram would be quite helpful: we'll be able to show exactly how framerate evolved during the benchmarking run. You'd think this is it, but there's one last topic that needs tackling ... and a rather troublesome topic it is.

content not found

Copyright 2022 ©

You may not use content, graphics, or code elements from this page without express written consent from

All logos are trademarks of their original owners. Used with permission.