Page 2 - AMD Radeon HD 6800 Series Architecture
AMD's Northern Islands family of GPUs is built on the 40nm fabrication process. Since TSMC couldn't deliver 32nm chips, the 6800 series will now serve as performance midrange cards to replace the position the 5700 series was in. This allows the 5700 series to continue at the $100 range, with the 6900 series taking on the high end enthusiast market. There will be three cards in the 6900 series and two cards in the Radeon 6800 series. The 6850 (Codenamed Barts Pro) commands a market price of around $160-$200 at press time, and the 6870 (Codenamed Barts XT) is priced slightly higher at $200-$250 retail when writing this review. This puts the card in relative competition with NVIDIA's GeForce GTX 460, despite having some price disparity between the two. When compared against its preceding generation, the AMD Radeon HD 6850 is faster than the 5830, with the 6870 slotted nicely between the 5850 and 5870. Initially, OEMs are free to make their custom, non-reference designs of the 6850 only. However, this has changed a few months back, and companies like Gigabyte are quick to jump on board with their proprietary 6870s. The Gigabyte Radeon HD 6870 1GB SOC we are examining today is a non-reference board that comes with the company's own WindForce 3X cooler in conjunction with a factory overclock for a price of approximately $220, or $20 than the 'standard' overclocked version with the same heatsink and fan. We will talk about why it commands the $20 premium -- but for now, you will be happy to know it comes with a 3-year warranty, and a custom cooler to boot. We will talk about the performance and noise of it later on in this review.
This second generation DirectX 11 card from AMD has a default core clock of 900MHz, and 256-bit memory running at an effective data rate of 4200MHz. Thanks to the increased clock speed over the previous generation cards, the Radeon 6870 at 900MHz can process 900 million polygons per second. As this is an evolution to the Evergreen family of graphics processors, AMD's Barts core continues to use AMD's VLIW5 design at the SPU level. The w, x, y, z, and t stream processors -- 5 in total -- execute instructions by working together with an array of general purpose plus a branch unit. On each clock cycle, the SPU can handle 4 32-bit floating point MAD, 4 24-bit integer MUL or ADD, and 1 32-bit floating point MAD for the SFU. To bring the size of GPU down, AMD dropped 64-bit floating point arithmetic from Barts. This is not going to affect gaming performance for the general consumer, but developers who are looking for GPU computing will need a different card.
Putting this all together, AMD managed to shave off 450 million transistors to hit a sweet spot in die size. Barts now has 1.7 billion transistors on chip, but it is still an astonishing number of almost two times the amount of a Sandy Bridge Intel Core i5/i7 processor at 995 million transistors. The Radeon 6850 has 32 render output units (ROPs), 960 shader processors, 12 Single Instruction, Multiple Data (SIMD) clusters, and 48 texture units. 4 blocks of 128K L2 cache and 4 64-bit memory controllers are attached to the ROPs for a 256-bit memory bus. Each SIMD cluster has 80 stream processors each, so you will get a total of 960 stream processors. Remaining the same as the Cypress core is 16KB of L1 texture cache, 8KB of L1 compute cache, and 4 texture units per SIMD. This brings you 1.5 teraflops of processing power at a reduced die size from 334mm^2 to 255mm^2, combined with a full gigabyte of 256-bit GDDR5 memory -- putting the Radeon 6850 a bit above the NVIDIA GeForce GTX 460 768MB. AMD's graphics processor retains a commendable 19W TDP at idle, and maxes out at 127W TDP under full throttle.
For reference, its bigger brother -- the Radeon 6870 Barts XT we are reviewing today -- is clocked 125MHz faster for a core speed of 900MHz. The memory bus width is the same at 256-bit, but has an effective data rate of 4200MHz. Our Gigabyte Radeon HD 6870 1GB SOC 50MHz faster than stock on both the core and memory (200MHz effective for the latter) than AMD's specifications. Besides clock speeds, all Radeon 6870s have 1120 shader processors rather than 960, 14 SIMD clusters rather than 12, and 56 texture units rather than 48 to bring the processor power up to 2.0 teraflops. In relation to the Radeon 6850, that is an additional 2 SIMD engines or 160 stream processors, as well as 8 additional texture units. Its idle TDP remains unchanged at 19W, but the load TDP is bumped to 150W. The Radeon 6870 competes against the NVIDIA GeForce GTX 460 1GB, although not directly.
Its features are quite similar to AMD's previous generation graphics cards. New stuff includes an update to DisplayPort 1.2 ready specifications and HDMI 1.4a compliance, and introduction to a new format of anti-aliasing called Morphological AA. Tagging along is the company's HD3D and UVD3 technology. Improved anisotropic filtering and tessellation performance due to better thread management and buffering compared to the 5800 series is also part of the package. AMD maintains its rotational invariance AF of the 5800 series, and claims the 6870 has twice the tessellation performance of the 5870. From AMD's internal testing results, we can see AMD has focused on tessellation performance at lower tessellation factors between 6 and 10 to back their claims. Between factors of 1-5 and 11-13, the tessellation rate compared to the Radeon 5870 is in the 1.5x range; and continues to decrease from there. AMD believes the lower factors are more crucial for gaming purposes.
I am not going to spend too much time going into specific details with regards to every new feature on the Radeon 6870. Basically, DisplayPort 1.2 goes hand in hand with AMD's Eyefinity initiative, since it is capable of efficiently driving up to six monitors without an independent clock generator for each display. Additionally, DisplayPort is a packet-based transmission medium; where devices are required to handle packets addressed to them. This is rounded into the term Multi-Stream Transport. DisplayPort 1.2 also doubles its preceding revisions bandwidth from 10.8Gbps to 21.6Gbps. By doubling this bandwidth, a single port can handle 2 2560x1600 monitors or 4 1920x1200 displays. 17.28Gbps of bandwidth is allocated for video, and audio can take as much as 48Mbps over the line. As far as HDMI 1.4a is concerned, is more TV focused rather than computer monitor centered. The entire 6800 series cards support 1080p 3D stereoscopy for movies, 720p for games and other content that require 60Hz per eye.
AMD's Unified Video Decoder 3 provides support for 3 new codecs, including MPEG-2, MVC, and DivX/XviD. Decoding is now completely offloaded to the GPU, compared to limited or no support in previous generation cards. UVD3 also supports Multiview Video Coding, which is an H.264 extension for 3D stereoscopy. Meanwhile, AMD's HD3D technology is touted as an open standard to go with the Radeon 6800 series cards. The company does not make its own 3D drivers, so it relies on third party middleware partners for driver wrappers, and like NVIDIA, it requires compatible active shutter glasses and monitors.