More Features of the RV770
The render back-ends in the RV770 GPU have been significantly enhanced over those in the RV670 as well, with the goal of improving anti-aliasing performance.
The improvements to the render back-ends resulted in double the peak rate for depth/stencil operations and double the AA fillrate for 32- and 64-bit color. Like the previous generation, they support both fixed function MSAA and programmable CFAA modes. In fact, with the Radeon HD 4800 series, AMD is introducing new Edge Detect CFAA filters that enhance AA image quality with 12x and 24x modes that have the same memory footprints as 4x and 8x MSAA.
Also new to the RV770 is a distributed memory controller architecture that's laid out around the perimeter of the chip, with individual controller segments situated adjacent to the primary bandwidth consumers. In comparison to the ring-bus memory controllers of previous ATI GPU, this new distributed design reduces latency, silicon area, and power consumption. In addition, the new memory controller also supports GDDR5 memory technology. GDDR5 is an evolution of GDDR3/4 technologies that allows for shorter traces, lower voltages, and increased bandwidth.
Along with the increased number of stream processors at work inside the RV770, AMD also incorporated a number of other enhancements. As we've mentioned, the RV770 is the first Teraflop GPU architecture but it also offers fast double precision (FP64) processing, thread generation, integer bit shift ops, and data sharing between threads in flight. Like we showed you with NVIDIA's latest GPUs, the sheer number crunching power of modern GPUs like the RV770 make them well suited to more than just playing games. There are a multitude of HPC applications where GPUs like the RV770 can yield huge performance improvements over current CPUs, and a number of desktop application that take advantage of the GPU are in the works as well, like video encoders for example.
To quickly test the number crunching abilities of the RV770, we enlisted the help of a small application dubbed "GPUQuant" that performs Black & Scholes or Monte Carlo calculations on either a CPU or GPU. The results show NVIDIA's GTX 200 series GPUs with a marked advantage in this test, but both AMD and NVIDIA GPUs offer huge gains over even a quad-core GPU.
In addition to all of the enhancements made to the stream processors, texture units, memory controller, and render back-ends, AMD is also unveiling a new video processing engine in the RV770, dubbed UVD 2. UVD-2 shares many features in common with the last-gen UVD engine in the RV670 and adds support for dual-stream playback hardware accelerated DVD upscaling.
Finally, we have a little teaser courtesy of AMD's presentation material. Eventually, AMD will be releasing a successor to the dual-GPU powered Radeon HD 3870 X2 based on the RV770 GPU. Currently, the product is code named R700, and according to the slide, a pair of them linked together in a four-GPU CrossFireX configuration can put up a 3DMark Vantage score of over 12.5K in Extreme mode - which happens to be a bit higher than three GTX 280s according to our benchmarks. Hopefully we can corroborate these numbers for ourselves at some point in the not too distant future.