A little more than three years ago, ATI released the Radeon 9700 Pro. The Canadian manufacturer surprised everyone and was the first to release a DirectX 9 chip. While NVIDIA was struggling to finalise the 130 nm GeForce FX, ATI used only a 150 nm fabrication process.
History seems to repeat itself, but this time the roles are reversed. If NVIDIA played it safe and chose a 110 nm process for the GeForce 7800 GTX (released three and a half months ago), ATI decided to be the first to use a 90 nm process for the R520.
In the end, the R520 required two additional revisions compared to usual product releases, which resulted in a four month delay. Now that this trouble period is behind them, ATI is ready to release the Radeon X1800.
New architecture, a quick lookThe new architecture for the 321 million transistors as introduced by ATI includes several optimisations. The memory controller has been reviewed to have a higher frequency and better use of available bandwidth. Memory is no longer accessed at 64 bits (4x64) at a time but at 32, and a memory ring bus has been developed.
This bus helps to simplify memory routing within the chip and increases memory frequency. Data read in memory is directly sent to the units requesting it via this bus and without going through the memory controller.
Cache was also modified as it no longer uses a Direct Mapped or N-Way Set Associative mapping type but a rather Fully Associative one. This increases the hit ratio (corresponds to a % of possibility for data requested by the CPU to be in the cache) at the expense of searching speed to know if a data is (or isn’t) in the cache. In theory, this last point isn’t a problem for ATI because of Ultra Threading, which we´ll now discuss.
As you probably guessed, ATI draws a parallel with Intel’s HyperThreading, which was introduced in the Pentium 4. The objective is the same: to best use the different pixel shader calculation units. If a unit is at rest and waiting for data to complete a calculation, it´s possible to put this calculation temporarily aside to process others.
In terms of functionalities, ATI catches up with NVIDIA. The accuracy of pixel shader calculations, previously “restricted” to 24 bits, is now increased to 32 bits similar to NVIDIA ever since the GeForce FX. Also, ATI now supports Pixel Shader version 3 and dynamic branching. They say that the impact on performance is lower than that of competing architectures.
High Dynamic Range is now also included as the Radeon X1800 is capable of processing blending for FP16 textures. This texture format using floating point calculation numbers is used in games such as Far Cry, Splinter Cell or Age Of Empire 3 for more realistic lighting. And what´s more is that if the GeForce doesn’t support anti aliasing when HDR is activated, the Radeon X1800 exceeds this limitation.
The number of calculation units is 16 (like the X800) for pixel shading and texturing. The vertex shading engine has been strengthened from 6 to 8 units. 3DMark05 will appreciate this improvement.