Just over a month ago AMD unveiled their new series of mainstream Radeon HD 6800 graphics cards based on Barts GPUs. And today, on December 15, 2010, they continue with their attack by releasing two top-end single-GPU solutions based on the Cayman graphics processor.
So far, there are only two cards in the series: Radeon HD 6950 and HD 6970. However, we won’t be surprised to see a third card with lower specs and a more affordable price, just as it was the case with the Radeon HD 5830. Moreover, a dual-processor flagship card based on the Cayman is scheduled for a Q1 2011 release as a replacement for the existing Radeon HD 5970.
AMD pits the Radeon HD 6950 and 6970 against Nvidia’s top-end GeForce GTX 570 and 580 cards. To make this even more exciting, the new products are expected to win not only through sheer performance but through competitive pricing and functional advantages. That’s an intriguing approach and we are going to discuss it in detail in this review.
AMD Radeon HD 6950 and HD 6970 2GB
Although the new Cayman processor is referred to as the HD 6xxx series, it is architecturally different from the HD 68xx as well as HD 5xxx solutions. The differences are both qualitative and quantitative, setting a milestone in the development of AMD’s new graphics architecture.
The most notable change is the transition from 5-way VLIW processors that debuted back in the R600 chip to 4-way VLIW ones.
The older 5-way VLIW processor contained five execution units, four of which could process only simple instructions such as addition and multiplication while the fifth one could also perform complex instructions like sine, cosine, square-root, etc. The new 4-way VLIW processor doesn’t have the complex-instruction unit as its functions have been transferred to the pair of simple units (numbers 3 and 4 in the slide). Thus, each 4-way VLIW processor is still capable of executing four simple or one complex instruction, which is the most typical scenario. It can only be inferior to the old 5-way VLIW processor in the rare situations when one complex and more than two simple instructions have to be performed simultaneously. This is unlikely in real-life applications whereas the higher efficiency of the new VLIW processors is going to increase the overall performance of the GPU at little cost.
The new GPU also features faster double-precision operations and an optimized mechanism of communication with system memory.
Another important improvement is the higher performance of the raster operators in different modes. The tessellation unit has also become much faster, according to AMD.
The rest of the innovations brought about by the Cayman processor are quantitative. The number of texture sampling and filtering units is increased from 80 to 96 while the number of VLIW processors is increased by 20%, from 320 (1600/5) in the Cypress to 384 (1536/4) in the Cayman.