4D is just the beginning: The official AMD Radeon HD 6900 architecture

AMD held an internal exchange meeting in Poland yesterday to explain to the local media the new architecture design of Radeon HD 6900 series graphics cards. Although we experienced the storm of leaks and deletions, we still saved key parts of them.

Interestingly, the lifting time for these slide shows was October 2010, proving that the Radeon HD 6900 series was indeed intended to be available last month but was postponed to next month for various reasons.

Although the Cayman Radeon HD 6900 and Barts Radeon HD 6800 are so close together, they are based on two very different core architectures. The Radeon HD 6800 continues the 5D-style stream processor design since R600, or VLIW5. Four simple plus a complex stream processor form a SIMD unit; Radeon HD 6900 is changed to a 4D design, also called VLIW4, each group. The SIMD unit consists of four equal mid-stream processors, plus a general-purpose raster unit and a branch unit.

In addition, the Cayman Radeon HD 6900 core has more SIMD arrays and texture units, dual graphics engines, eighth-generation tessellation, improved rendering backends, new GPU computing power, and enhanced quality anti-aliasing (EQAA) High-speed 256-bit GDDR5 memory interface with bandwidth exceeding 5Gbps, etc. We will explain each one below.

The VLIW4 4D architecture uses a four-way concurrency design. All stream processors have the same integer and floating-point execution capabilities (no longer a T-Unit), but two of them have additional special features.

AMD claims that the VLIW4 architecture can reduce core area by approximately 10% compared to VLIW5, while simplifying scheduling and raster management, and the logic core has also been re-used extensively.

Although there are no such stream processors on the Cayman Radeon HD 6900 core, there are 3840 Radeon HD 6990s based on the previous news, so the top-level single-core Radeon HD 6970 is naturally 1920. 30 SIMD array engines.

In terms of front-end design, the Cayman Radeon HD 6900 core uses two graphics engines and can handle two primitives per clock cycle, with block-based load balancing, and conversion and hidden face elimination rates doubled.

At the same time, two rasterizer units can process up to 32 pixels per clock cycle.

And two eighth-generation tessellation units (Barts Radeon HD 6800 is the seventh generation) that support off-chip buffering for higher tessellation levels to improve performance, up to three times more than the Radeon HD 5870.

The tessellation technology has become a hot spot for DX11 because of the company's publicity. AMD has even evolved it to the eighth generation and claimed that its practice is correct, including focusing on the most effective tessellation model. Adapt to surface subdivision and so on, can achieve the balance of performance and picture quality.

With different surface subdivision factors, Radeon HD 6970's surface subdivision performance can reach 1.5-3.0 times that of Radeon HD 5870.

Improvements to the renderer backend include a combination of write operations, a 2-fold increase in 16-bit integer (unorm/snorm) operations, and 2-4 times faster 32-bit floating point (single/double precision) operation.

GPU computing enhancements include asynchronous allocation (synchronous execution of multiple compute cores/each kernel has its own command queue and protected virtual addressable domain), two bidirectional DMA engines (faster system memory reads and writes), and shader reads Merge, LDS (local data sharing) direct picking, improved flow control, faster double precision operations (single precision one quarter).

Enhanced Quality Anti-Aliasing (EQAA) is a new multi-sample anti-aliasing (MSAA) with up to 16 sampling points per pixel. The number of colors and sampling points can be controlled independently, and the sampling mode and filters can be configured by themselves.

AMD claims that the technology can bring better picture quality based on the same capacity of video memory, and is compatible with adaptive AA, super-sampled anti-aliasing (SSAA), and form anti-aliasing (MLAA).

As the area of ​​the core increases, power consumption control becomes more important. The Radeon HD 6900 integrates a power-consumption processor, monitors power consumption in real time and dynamically adjusts every clock cycle, and can directly process GPU core power consumption, no longer bypasses frequency and voltage adjustment, and also supports AMD. OverDrive tool.

The final specification of Radeon HD 6970 has not yet been explicitly announced. The number of triangles processed per clock cycle is two (Radeon HD 5870 one); the Z/Stencil ROP units, Color ROP units are still 128 and 32 respectively; the SIMD engine, texture unit only say more than 20 and 80 respectively, guessing Should be 30 and 120; memory bandwidth is greater than 160GB/s, indicating that the memory equivalent frequency will indeed exceed 5Gbps.