X-bit labs

First in DirectX 10.1: ATI Radeon HD 3850/3870 (page 5)

Category: Video

by Alexey Stepin , Yaroslav Lyssenko, Anton Shilov

[ 01/25/2008 | 10:31 AM ]

Texture and Raster Processors

The texture and raster processors of the Radeon HD 3800 do not differ from those of the Radeon HD 2900 architecturally. They are complex devices we can only roughly compare with traditional TMUs and ROPs. The RV670 incorporates four large texture processors each of which contains:

• 8 texture address processors
• 20 texture samplers
• 4 texture filter units

As you see, the RV670 should have no problems with texture sampling but each processor cannot filter more than four texture samples per clock cycle, 16 samples in total. Clearly, it is the filtering of textures that is the bottleneck in the texture processors of the Radeon HD 2000/3000. There is only one filter unit per each two address units, so there is no talking about free tri-linear or anisotropic filtering. It is here that the RV670, like the R600, is inferior to Nvidia’s G92 and, especially, G80 GPU with its two filter units per each address unit. There are some improvements, though, but they are on a deeper level – we’ll discuss them in the DirectX 10.1 related section below.

Each raster processor of the RV670 contains:

• 4 Alpha/Fog units
• 8 units working with Z- and Stencil-buffers
• 4 Blend units
• 16 multi-sampling processing units

So, it is roughly equivalent to 4 classic ROPs and can process four pixels per clock cycle for a total of 16 as there are four such processors in the chip. When working with the Z-buffer, i.e. with pixels that don’t contain color information, the performance is doubled for a total of 32 pixels per clock cycle. Like the texture modules, the RV670’s raster processors have been redesigned to support DirectX 10.1 capabilities.