R600 details revealed
ATI's new graphic processing unit goes by the codename "R600" and is ATI's second generation unified shader architecture. A wide array of programmable shader units are used to do either pixel, vertex or geometry shading which looks very much like NVIDIA's G80 GPU though on a closer look there is noticeable difference between each GPU architecture.
One of the mayor architectural components of the R600 is the Ultra-Threaded Dispatch Processor. This part of the GPU makes sure that the unified stream processors do the right calculations on the right piece of data. You can see it as the team leader, in this case his goal is to keep the efficiency high by filling up as much stream processors as possible so that the data is being processed as fast as possible.
The stream processors are divided over 4 SIMD (Single Instruction, Multiple Data) arrays each consisting out of 16 Superscalar unified shader units, which on their turn exist out of 5 stream processors. This means that the R600 has a total amount of 64 Superscalar shader units or 320 stream processors!
Each Superscalar unit exists out of 4 normal stream processors plus special stream processors which supports more different kinds of instructions. Besides that there is also a branch execution unit which brings the total amount of instruction per clock cycle to 6 per superscalar unit. Compared to the G80 which has only 128 stream processor clocked at 1350MHz and able of doing only 1 instruction per clock cycle, this means that the R600 is definitely the winner when it comes down to raw processing power. The downside of its architecture is the complexity to get everything going well, so it's up to the driver programmers now to provide the end-users with as much performance as possible.
The GPU itself exists out of 700 million transistors and is build on a 80nm HS fabrication process. The R600 supports FP16 texture filtering at full-speed 128-bit HDR rendering and is Direct 3D v10.0 and OpenGL 2.0 compatible. One of the improvements over the past generation GPU's is the support for up to 8xMSAA (multi-sample anti-aliasing) using programmable sample grids which makes it possible to adjust the anti aliasing setting by driver updates. Another new AA mode is also being introduced and goes by the name: Custum Filter Anti-Aliasing. With CFAA the filtering method is no longer based on a box, this time it looks at the nearby pixels around the particular pixel that is being processed before the pixel itself is being adjusted in color and the image being anti-aliased. As this filtering method is being processed by the shader processors it also brings better image quality because of the flexibility to program the filtering method, though the downside here is that less shader processors will be available for doing pixel, vertex or geometry calculations.
ATI's familiar ring bus technology is being used to connect the DRAM chips with the R600 processors. On our HD 2900 XT we found the above GDDR3 chips manufactured by Hynix. Although the HD 2900 XT has a memory clock speed of 828MHz in 3D mode, Hynix official rated those chips for up to 1000MHz DDR and so there should be quite some overclocking headroom left in the HD 2900 XT's memory.
On ATI's reference HD 2900 XT sample we also found the ATI Theater 200 multimedia processor. The second generation Theater chips include a wide range of video input formats and digital audio ports which makes the Theater 200 an ideal solution for a variety of multimedia applications including: Set-Top Boxes
PC & Internet Televisions
Entertainment/Home Theater PCs
More info can be found on ATI's website.