AMD's HD 6970 "Cayman" Architecture Unveiled

Courtesy of NGOHQ.com, we gained insight into the way how AMD refined the Evergreen architecture and how Cayman should be competitive against nVidia's own architectural refinement - Fermi "1.2".


AMD Radeon HD 6900 Series Slide with the expired NDA

On the first slide, we see that the NDA [Non-Disclosure Agreement] date was set for yesterday, Monday November 22, 2010 at 12:01AM EST but as we all know, the company had to delay the launch to the week starting on December 13, 2010.

The Cayman Islands on which you can't establish an off-shore bank account

What is Cayman? Two Graphics Engines with fifteen 64-core Clusters each for a grand total of 1920 cores

In the second slide, you can see that Cayman GPU actually features a Dual graphics engine, essentially, you can compare Cayman to two Barts GPUs combined into one. You still have the same 256-bit memory controller known from Evergreen parts [Cypress] and Barts, the affordable Northern Islands part [Radeon HD 6800 Series], but now the memory controller is loaded from two independent graphics engines.


Dual Graphics Engines are consisted out of Geometry and Vertex Assemblers, Tessellator unit and Rasterizer/Z operations feeding into Shader Engine

Both Graphics engines receive orders from one Command Processor, but they're aggressively performing load balancing using the tile-based method, much akin to the way how certain multi-GPU methods work. Each Graphics Engine is consisted out of Rasterizer [ROP] and aggressive Hierarchical Z from one side and Geometry Assembler, Vertex Assembler and Tessellator from another.

Do bear in mind that even though AMD claims to have an 8th generation tessellator processor, game developers we talk to often are mentioning Northern Islands will be AMD's second generation of usable tessellators. According to an unknown AAA game engine architect, the problems with Tessellation was that in order for game engines to adopt Tessellation, Tessellators needed to have five key features and lack of a single feature rendered Tessellation unusable. With DirectX 11 coming onto the scene, Microsoft finally mandated all key features of Tessellation unit into its own API and both AMD and nVidia had no other but to follow suit. The company claims the tessellation performance improved on average of 200%, with the sweet spot being 7-9x Tessellation Points, where improvement is measured up to 300% over HD 5870.

Smart/Simple VLIW5 concept finally gave way to a more efficient, "all-smart" approach called VLIW4

Going with a Dual Graphics engine is a natural consequence of the way how AMD increased efficiency of their cores. For starters, for the first time since 2004 [Xbox 360/Xenos GPU, R520 "Fudo"] that AMD got rid of VLIW5 and simplified its execution core into VLIW4: four units now have equal capabilities, thus removing the VLIW5's "one smart, four simple unit" concept. Now, all four units are "smart" ones, capable of executing either Floating-Point or Integer operations. Special functions are executed by combining three to four units together.


Power Efficiency and Overclocking: Now done through AMD Overdrive app

For those overclockers among us, AMD is finally syncing its excellent Overdrive utility with the power consumption of the graphics cards. AMD Radeon HD 6900 Series cards feature integrated power control processor which monitors the power draw of the GPU, but also enables user to directly control it. No more voltage tweaks, just clearly define what is the maximum power draw by the GPU and the power supply will deliver it to appropriate component on the graphics PCB.


EQAA is compatible with all AMD-supported methods of AA: adaptive AA, multi-sampling, super-sampling, morphological AA

Last but not least, Northern Islands bring a new Anti-Aliasing mode called Enhanced Quality Anti-Aliasing, i.e. EQAA. Just like nVidia's approach to 32x Anti-Aliasing, AMD decided to utilize ample computational performance of Cayman GPU and created a mode that enhances existing AA algorithms, combining the conventional Color Sample Location [up to 8x] with Coverage Sample Location, i.e. up to eight computationally sampled locations per pixel for a grand total of 16 samples per single pixel.



In case of nVidia's 32x AA, we were greeted with an improved Anti-Aliasing effect over the non-computational modes, while we expect that AMD could enable the same level of quality with 16x EQAA.

Conclusion before the conclusion
All in all, Northern Islands seem like an worthy architectural upgrade over the Evergreen series. Given that AMD was stood up by TSMC on the 28nm process dropping back all the way into late 2011, there was no choice but to spend nine months on refining the Evergreen architecture and improve the shortcomings. Just like nVidia, major issue with the Cayman and other parts will be yields coming out of troubled 40nm process. Regardless of what both AMD, nVidia and TSMC are saying [political correctness ahoy], the current state of the market and availability of the parts speaks for themselves.

Key question that AMD needs to answer though, is something they haven't been asked since Radeon HD 4800 Series came along: Can Radeon HD 6900 Series keep up with nVidia's refined Fermi, i.e. GeForce GTX 580, the upcoming GTX 560 and 570, as well as the dual-board GF110 [GTX 595] and GF114 [GTX 565] based parts?