David Kanter (dkanter@realworldtech.com) on 4/12/11 wrote:
---------------------------
>As many of you may have noticed, we've had a profusion of new articles at RWT in
>the last few months. I have been writing one short article per week, for the last
>5 weeks, while still working on several longer pieces (Poulson, z196, cough cough).

z196? Do tell.


>Without further ado, here's the deal on this weeks article:
>
>Modern graphics processors are incredibly complex, but understanding their performance
>is essential, as they become an increasingly important component of computer systems.
>In this report, we use a set of benchmark results to build accurate performance
>models for AMD and Nvidia GPUs. We verify that our model can predict performance
>within roughly 6-8% for many desktop graphics cards and show how Nvidia’s microarchitecture
>and drivers achieve roughly 2X higher utilization than AMD’s VLIW5 design.

The problem with benchmarks and graphics is that it is very easy to run into some sort of bottleneck. The data points were generated using a 1280 x 1024 resolution to minimize the affect of fill rate and memory bandwidth. Fair point until you start to run into CPU bottlenecks. This may only appear with the high end GPU's but it will skew the results. Similarly bench marking at 1920 x 1080 would produce subpar results for low end GPU's for the reasons previously mentioned. I feel that the best way is basically to run the analysis several times at different resolutions. Perhaps a 3D plot of just nVidia GPU's with the three axis labeled: GFLOP, 3Dmark and megapixel?

Another thing I'd like to have seen would be some love for older graphics cards (Geforce 8800GTX, Radeon 3870 ect.). My real curiosity here with the internal changes between the Radeon 2000, 3000 and 400 lines. While AMD didn't fundamentally chance their VLIW5 shader core during these generations, they modified many other characteristics. The internal ring memory bus of the Radeon 2000 was dropped, the number of TMU's and ROP's increased in efficiency, especially when dealing with anti-aliasing. If the analysis holds up for the modern cards, it should be able to accurately predict the results of the older cards: you don't necessarily have to wait for something new.