Architecture

AMD's dual-GPU graphics cards in recent times have carried a simple design methodology: There are two graphics processors, each with its own memory and voltage regulation circuitry. One of the two gets to be connected to the displays. Both GPUs then have PCI-Express x16 connections to a PCI-Express bridge chip, which broadcasts data from the host system, to which it is directly connected over the card's PCI-Express x16 interface. The two GPUs have a CrossFire connection (through the bridge), and can connect to up to two more GPUs of its kind over external CrossFire (in which there are multiple graphics cards connected using CrossFire cables). The card's internal CrossFire between its GPUs is enabled by default, not requiring the user to do anything beyond installing the accelerator like any other. The internal CrossFire is neutral to the system it is installed on, and hence it doesn't matter which make the system's chipset is. Have PCI-Express x16, will work.



With the Radeon HD 5970, it's no different. Each GPU is based on AMD's TeraScale 2 Architecture that offers over 2 TFLOPs of processing power. Each GPU has 1600 stream processors, arranged into 20 SIMD engines, each with 16 thread processors that have 5 processing cores each. There are 80 texture units per GPU, and 32 raster operation units. Built on the 40 nm silicon process by TSMC, the GPUs employ new energy-saving performance profiles that are found to offer some of the lowest power draws in idle, and some of the highest performance per Watt, when on the job. Each GPU uses a 256-bit wide GDDR5 memory interface, which provides 128 GB/s of memory bandwidth (256 GB/s for the entire card). The slightly improved PCI-Express bridge chip now supports the new PCI-Express 2.1 standard. According to AMD the new PCI-E 2.0 bridge chip has so much bandwidth that the mythical sideport from HD 4870 X2 can no longer be useful - not even in theory.