R520 Architecture

Pipes/Shader/units and the like:

First of all, let’s talk about the new X1800 3D architecture. It’s a 16 Pixel Shader processor chip, featuring an Ultra – Threading Dispatch processor and 4 Shader cores. There are also 8 Vertex Shader processors, 16 Texture Address units, 16 Texture units and 16 Render Back-to-Ends. You have probably noticed the absence of the word “Pipeline” from all of that, and it is intentional – ATI believes that it’s no longer useful to talk about how many pipelines a part has, since its efficiency is what determines its performance, not the number of pipelines. So, if we really wanted to describe the X1800 in “old” terms, we would say that it is a 16 pipeline part. That should put some of the 32/24/32 with 8 broken/32 with 16 broken pipeline myths that have been circulating the past months to rest once and for all. There is also the X1600 series, which has 12 Pixel shader processors, and the X1300 series with only 4. As mentioned already, the new series of cards is fully Shader Model 3.0 compliant, and in the following paragraphs we will see the features ATI believes are the key to this new architecture.

One of the buzz words that everyone’s been hearing when first Shader Model 3.0 hardware was announced was Dynamic Flow Control – ATI fully supports it, with Branching (IF…THEN…ELSE statements), Looping and Subroutines. That allows the X1000 series of cards to execute different paths through the same shader on adjacent pixels. This architecture also gives the developers many opportunities for optimizations, since early parts of a shader which that don’t need to be executed can be skipped on the fly, and plus, developers can combine many related shaders into one uber shader; a method that has the advantage of avoiding state change overhead (read: speed penalty) when going from one shader to the next. ATI insists that its dynamic flow control implementation is not a “tick-box” technology, but it actually solves real problems and helps with performance and detail management.

For dynamic flow control acceleration, ATI uses a new technique which is called “Ultra-Threading”. As we said, the key word that characterises every aspect of the new X1000 series is “efficiency”. What ATI wanted to do with this generation of graphics cards was to always keep the GPU busy, no matter what; and that gives an actual speed increase, since the GPU uses various techniques so as not to sit idle, waiting for things to happen (which is one of the most typical reasons that slows down a GPU these days). So what is Ultra – Threading? ATI decided to use hundreds of simultaneous threads across multiple cores. Each thread can perform up to 6 different shader instructions on 4 pixels per clock cycle, and the analogy is 16 pixels per thread (on the Radeon X1800). The reason for using many threads is really simple, if you stop and think about it. When you have many threads, you can essentially do many things at once, since obviously the large number of threads promotes parallelism. A single thread architecture obviously has a disadvantage here, since if you want to do several things, you have to do them one at a time (which obviously takes much more time than doing these all at once). You also have Fast Branch Execution thrown into the mix, with dedicated units that handle flow control without giving any overhead to the ALU (Arithmetic Logical Unit) and large Multi Ported registry arrays, which results in fast thread switching.