AMD Radeon HD 6850 and Radeon HD 6870 review
Author: Luka Rakamaric
Date: 10 Nov 2010

This year we won’t see a new die production process for ATI cards, as was the case last year when they introduced the 5000 series. Instead, we will have to make do with a just a refresh of the current 40 nm production process. Originally, AMD’s plan was to produce Barts in a smaller, 32nm production process. However, TSMC just couldn’t make it viable, and for two reasons. First, as AMD has experienced last year, their 40nm production process was quite problematic, resulting in low yields. Instead of working on three fronts, fixing 40 nm, and designing both 32 and 28 processes, TSMC decided to abandon the 32 nm and move directly to 28. However, not in time for Barts, so AMD was left with only 40 nm at its disposal. That kind of sealed the design of Barts, and so we are left with more of an evolution product rather than a new design.

The GPU
AMD decided to go with a slightly different approach than before. In introducing the midrange part first, it has abandoned the strategy of producing cut down versions of high end GPUs as their lower tier products. The whole new series of GPUs is called Northern Islands, which will be the basis for series 6000 cards. Today’s GPU is called Barts, in its Pro and XT variants. As we have mentioned before, Barts is an evolution of Cypress, but has been moved from high end to midrange. In fact, it is smaller than Cypress, both in SP count, but even more in die size and transistor count.

The full Cypress GPU had 1600 Stream Processors and a full Barts GPU has ‘only’ 1120. This has brought down the transistor count from 2.15 to 1.7 billion, and the die size from 334 to 255 mm2. In order to get this kind of drop, AMD had to also sacrifice some of the functionalities that existed on Cypress, most notably the FP64 operations. The architecture remained the same, with AMD’s (or ATI’s) classic ‘five stage’ pipeline. Four of the SPs are simple units, and the fifth is a special function unit that can also function as the previous four. This of course means that in certain situations that arise quite often, not all of the GPU is utilized to the fullest. However, AMD’s GPU’s have functioned like this for five generations now, and more often than not, this approach has produced excellent GPUs and cards. Just like in Cypress, 80 SPs form a SIMD, which in turn has 24 KB of cache attached to it, with 4 texture units.

Despite the drop in SPs, 32 ROPs still remain, as does the 256 bit memory bus. The memory controller also received a change, as it is more similar to Redwood’s than Cypress’. This is why memory frequency dropped from1.2 GHz to 1.05 GHz, but it also saved AMD a lot of transistors, which made the GPU easier and cheaper to produce and made it more power efficient. AMD also put a new tessellation unit in Barts, which will help to close the gap to NVIDIA that existed in the 5000 series. From AMD’s graphs, it should have a bigger performance increase at lower tessellation factors, with the performance approaching that of the 5870 as the factor grows.

The GPU also has the new UVD3 video decoding engine. It adds support for MPEG-4 ASP, MPEG2 and MVC. The first two are not that big of a deal as they are not that demanding on today’s systems. MVC or Multiview Video Coding is an extension of the current H.264 to allow for the 3D stereoscopic processing on Bluray 3D. AMD also included a new AA mode called morphological AA. It is already known on consoles, and it is, unlike other AA modes, a post-processing filter. It is applied after the image has been rendered and as such doesn’t bring much of a performance penalty, but will not provide such quality as supersampling AA or even multisample AA.