AMD Radeon HD 6900 Series: Right Performance, Right Features, Right Price
A New Architecture From AMD
The architecture of the Cayman GPU is fairly new, with quite a few changes from that of the Radeon HD 5000 series and Radeon HD 6800 series. The chip features two "front-end" graphics engines, with two thread dispatch processors, two geometry tessellation units, and two vertex processing engines. This design should dramatically improve performance in geometry-intensive applications, particularly those that have a lot of tessellation. (Tessellation is a DirectX 11 feature that breaks up large polygons into lots of smaller polygons, to add detail or to smooth the edges of 3D models.) Here's a breakdown of some of the new features in the Cayman GPU found on the Radeon HD 6900 cards.
PowerTune technology: CPUs from Intel and AMD have a "turbo" feature that can crank up the clock speed a bit if the chip happens to be running beneath its rated thermal threshold. The new PowerTune feature of the Radeon HD 6900 cards sort of works like that, but in reverse: The core clock speed is set very high, and if the chip starts to run too hot, PowerTune lowers the clock speed in small steps until it's back at a safe level. Most graphics cards have to set the clock speed conservatively for the worst-case scenario of an application that strains the GPU and makes it run especially hot. You'll find a new slider in the Catalyst Control Center that lets you tweak the maximum thermal threshold at which PowerTune will kick in and slow down the card, from -20 percent to +20 percent. This will be an interesting function for overclocking enthusiasts to play with, but it has even bigger implications for the future, when this architecture eventually finds its way into AMD's "APUs" (combination CPU and GPU).
New antialiasing modes: In the Radeon HD 6800 series, AMD introduced a new antialiasing mode called Morphological AA. This post-processing effect antialiases everything on the screen, but can make the image appear slightly softer as a result. In the 6900 series, AMD has added Enhanced Quality AA. It's functionally similar to Nvidia's "CSAA" modes in that it adds a set of coverage samples to the usual number of color samples. This type of antialiasing can improve image quality with a pretty small impact on performance.
New VLIW4 shader architecture: The last several generations of AMD graphics chips used a VLIW5 architecture (that is, a five-way Very Large Instruction Word vector unit in each stream processing unit). One of the vector units in each set of five could handle special functions such as transcendental operations. Now, each vector unit is a four-way unit, with no special-function unit; if a transcendental operation is called, it instead occupies three of the four units. If all that sounds like a bunch of gobbledygook to you, you're not alone. The gist of it is that it should be easier for the GPU's scheduling hardware to keep all the math-processing units busy, providing for better performance in the same GPU area.
Enhanced geometry features: The Radeon HD 6900 series cards feature two graphics engines (the front-end processing of the graphics pipeline). That means two tessellation units, two geometry and vertex processing units, and two rasterizers. To make a long story short, the chip should be able to process a lot more geometry than previous AMD GPUs could. We still think that the high-end cards based on Nvidia's Fermi architecture, which also boast multiple geometry processing units, should offer somewhat higher overall geometry performance.
GPU computing enhancements: Along with reworking all those shader units from VLIW5 to VLIW4, AMD has made a number of other enhancements that should help with performance in applications that use the GPU for general computational tasks (such as physics, image post-processing, or video transcoding). The tweaks include improved flow control, dual direct memory access (DMA) units for faster memory reads and writes, and the coalescing of shader read operations. Most interesting is "asynchronous dispatch," or the ability of the GPU to work on multiple compute kernels simultaneously. In other words, you could ask the GPU to do some graphics work, some physics work, and some video transcoding, and the chip could split up those loads to its various parts to work on simultaneously, instead of switching back and forth among them. This function has limited utility today, but could be important in the future as the GPU becomes more of a shared system resource.
Beyond those changes, the new Radeon HD 6900 cards share the improvements that AMD made in the 6800 series: the UVD3 video engine, DisplayPort 1.2 and HDMI 1.4a support, per-display color correction, and HD3D (AMD's branding for 3D-display support).
Next: When the rubber hits the road