AMD is taking on a new strategy with video cards. They want to give you the “perfect gaming solution for $200-$300.” That is certainly an admirable goal. We know that NVIDIA’s GeForce GTX 280 offers a great gameplay experience, but it also costs over $600. AMD’s strategy is a scalable one; think smaller GPUs in parallel. In fact, the new ATI Radeon HD 4800 series GPUs are only 16mm across versus NVIDIA’s 24mm across on its GTX 200 series GPU. The Radeon HD 4800 series GPUs are indeed small; they are literally the size of a dime.
AMD has made significant improvements in the architecture, in every component. The AMD ATI Radeon HD 4800 series is based on a 55nm process with 956 Million transistors. They have bumped the stream processor count up to 800 (versus 320 on the HD 3000 series.) They have also bumped up the texture units to 40 from 16 in the last generation. However, the ROP count stays the same at 16, though there have been improvements made to the ROPs to increase performance. Before anyone asks, yes, they have fixed the AA performance.
The stream processors have undergone many changes; they are now laid out as 10 SIMD cores, each with 80 32-bit stream processors (800 total). They also now can do fast double precision processing with integer big shift operations for all units. There are 80 scalar stream processors per unit with 16KB of local data cache that is shared. Each unit has its own control logic and each has 4 dedicated texture units plus L1 cache. Tessellation performance has also been improved. Of course, DX 10.1 is supported as well.
The texture units have also received quite a bit of work; there is double the texture cache bandwidth of the HD 3000 series. The entire cache system has been reworked to improve performance. The ROPs have also been improved to allow better AA performance among other things. There is also a new AA mode being introduced called CFAA which stands for Custom Filter AA. There will be a drop down box in the control panel where you can select the option called “Edge Detect” which will enable the CFAA modes. “Box Filter” is the regular AA modes.
The way CFAA works is really rather simple, it uses the stream processors to do edge detect AA to provide better AA image quality, up to the equivalent of 24X AA. The great thing about it is that it does not use or need a larger memory footprint for doing this as it does not tax the framebuffer. It does however tax the stream processors, so in shader intensive games it may not be possible unless you have cycles to spare. In the games we are testing today we did not find it playable in any of them with the highest playable settings. We will be doing further investigating of performance and image quality of CFAA modes in a later evaluation and dive into more detail then.
There have also been other various improvements, such as memory controller optimizations to support the new GDDR5 memory modules. GDDR5 offers double the bandwidth of GDDR3/4 basically, and allows AMD to achieve high bandwidth on a 256-bit memory bus.
HD 4870:

HD 4850:




News