ASUS EAH 6850 Direct Cu and EAH 6870 Print E-mail
Written by Michael Schuette   
Nov 28, 2010 at 07:41 PM

It is this time of the year where everybody comes out with the next model. Maybe it is a bit late for that in the car industry but computers and computer components still abide by slightly different rules. And the manufacturers are not accepting bail-outs either, or maybe they were never offered any… who knows.

There has been a lot of turmoil lately in the graphics adapter scene. nVidia’s introduction of the Fermi architecture has resulted in a mixed reception for their GF 100 product that arguably was the most powerful card at its time, yet, as we pointed out in our review of the GFGTX 480, a bit overdesigned in some areas and a bit anemic in others, particularly, the memory interface.

Along came the GF 104, a bit leaner, a bit trimmer, a bit more balanced and a lot less expensive for almost the same performance, at least in the current gaming landscape, thereby establishing at least some kind of competition for AMD’s omnipresence and omnipotence in the bread and butter market. Naturally, AMD had to counter with a new lineup of GPUs, code name Northern Islands as opposed to the Evergreen series that was so dominant during the last year.

Enter Barts, the new AMD GPU manufactured at TSMC at the – finally viable - 40 nm process and designated to maintain the lead in the $150.- to $250.- market segment while replacing Cypress. Interestingly, the Juniper GPU powering the Radeon HD 5700 series will live on throughout the next generation cycle.

To recap, Cypress had a total of 20 SIMD arrays, each having 16 execution units that were 5-ALUs wide to manage a total of 1600 arithmetic logic units dedicated to crunch the various mathematical processes making up the plethora of programmable shaders under the DX11 umbrella. Similar as in the case of nVidia’s GF 100, the processing appeared to be a bit on the overkill side compared to the memory interface, at least, that’s what some of our benchmarks of “special edition” releases by the channel partners suggested to us. I guess, we were not the only ones to come to that conclusion, in fact, it looks as if Eric Demers and his cohorts did at least some similar math and came to the according conclusions.

The result is Barts where the focus appears to have been on the 80/20 rule, that is, deliver 80% of performance at 20% of the cost and then scale up the design again to beat the predecessor in every which way at 60% of its cost. In more detail, we are looking at 14 SIMD arrays, without really changing much on the internal organization for an end result of 1120 shader execution units. The result is a dramatically reduced die size, going back from 334 mm2 to 255 mm2 with the transistor budget scaled back from 2.15 billion to 1.7 billion.

One thing you may notice is the dual listing of rasterizers. This has not really changed from the Cypress design and, like in the former it comes down to one triangle per cycle per GPU. That is, the geometry throughput is a 1:1 function of the clock rate, in contrast to nVidia’s offerings that are pushing 2 or 4 triangles per clock in the GF104 and 100, respectively.

Along with the die size and transistor count reduction came a power reduction. That is, the load power and the idle power of the R5850 and the R6870 are identical, however, the same power envelope allows to crank up the GPU core clock from 725 MHz to a whopping 900 MHz. As a result, whenever the cards are idle, the power consumption of the new design drops almost by 33%, or, conversely the older, Cypress-based R5850 had a 50% higher idle power consumption than the new Barts-based R6870. The wonderful world of math allows us to use either metric depending on what we want to show.

Discuss this article on our forums

Last Updated ( Dec 20, 2010 at 03:35 AM )
Next Article>