Architecture and Technical Characteristics
The modular concept that ATI Technologies sticks to helps create GPUs with varying performance by employing different combinations of functional subunits.
The 120 shader processors of the RV630 chip are joined into blocks of 5 ALUs, and one processor is capable of executing instructions like SIN, COS, LOG, etc. The core incorporates a total of 24 such blocks organized into 3 large computing modules with 8 blocks or 40 ALUs in each. It means we can expect a higher computing performance from the RV630 than from the Nvidia G84 that has only 32 shader processors. And this may be the decisive factor in games that focus on special effects with complex math1ematics as well as in DirectX 10 games.
On the other hand, the RV630 is far inferior to Nvidia’s GPU in terms of texture-mapping and rasterization performance as it has only two large texture processors roughly equivalent to 8 TMUs and one render back-end equivalent to 4 ordinary raster operators. This is going to have a negative effect on the performance of the Radeon HD 2600 when processing large amounts of high-resolution textures.
AMD must have chosen this number of texture-mapping and rasterization subunits because the chip was already too complex. Moreover, the texture cache is reduced from 256 to 128KB and the ring-bus width decreased from 512 to 256 bits in the RV630 chip.
As you see, even configured like this, the RV630 has 101 million transistors more than the Nvidia G84 has. If the RV630 had more texture and raster processors, the chip just wouldn’t have fitted into the desired price category.
There is one more important difference from the R600: the RV630 and RV610 chips incorporate a hardware video-processor based on the ATI Xilleon architecture and capable of decoding HD video in H.264 and VC-1 formats off the CPU. For comparison, the G84 and G86 offer full hardware decoding only for H.264 while the first stage of VC-1 decoding is done on the CPU.
The existing versions of ATI Radeon HD 2600 XT GDDR4 differ in memory type and frequency. The low memory frequency of the Radeon HD 2600 XT GDDR3 must be due to the manufacturer’s desire to cut the cost by installing slower chips although there are no fundamental obstacle to clocking such memory at 1000 (2000) MHz or higher. Anyway, the price difference isn’t big at only $20. The Radeon HD 2600 XT GDDR3 shouldn’t be far slower than the Radeon HD 2600 XT GDDR4 because each card is going to be limited by the performance of their TMUs and ROPs rather than by the memory subsystem. We’ll check this out in the Tests section soon.
We’ll discuss the design of the Radeon HD 2600 XT using its GDDR4 version.