AMD Radeon HD 6870 and 6850 Review - Barts Architecture Refresh
A new architecture we are all very familiar with
It has been a good year from the AMD GPU team. Since the release of the Radeon HD 5000 series of cards back in September of 2009 there has been a resurgence of noticeable size in the amount of AMD/ATI GPU users and we ourselves, among countless other hardware enthusiasts, touted the Evergreen architecture and the corresponding graphics cards as outstanding. For a good a six months AMD enjoyed an exclusive on DX11-class hardware and used it, as well as killer features like Eyefinity, to woo gamers back to the red team.
This lead netted AMD a 90% share of the DX11 market and allowed AMD to output 25 million GPUs across their spectrum of cards.
And a spectrum, it is. The Radeon HD 5000 series consists of 11 cards in total ranging from the HD 5970 dual-GPU offering to the tiny, low powered, HTPC-ready HD 5450.
But now it is time to move past the 5000 series and discuss what is next, what is available starting today: the Radeon HD 6800 series.
The "Barts" Architecture - Radeon HD 6800 series
The Barts chip, known officially known as the HD 6800 series of cards, is based almost completely on the Evergreen architecture that we detailed more than a year ago in our HD 5000 series launch article. But while this launch might be viewed as a rebranding by some people, there were some notable changes in the architecture and design of the chip that allow AMD to offer improved performance and better performance per watt/area.
Barts will also be bringing the Evergreen architecture seen in the HD 5800 series of card down into the sub-$200 market - no doubt a direct reaction to NVIDIA's GeForce GTX 460 cards that have been very successful in their short life on shelves. Later in the year we will see the release of future architectures that much more unique in the Cayman and Antilles product lines. We'll have to leave you with that tease for now and touch again on both of those items later.
So if Barts is based on the currently available HD 5000 Evergreen architecture, what was AMD able to do to improve performance and compel us to accept the new HD 6800 cards? AMD was definitely pushing the fact that the Barts design has improved performance per die area - and while this is true their claim of a "35% improvement" is a bit of a stretch. In that metric they were comparing the new HD 6870 GPU (which is a fully enabled chip) to a Radeon HD 5850 (which is an HD 5870 with part of the chip disabled) which is not a completely legitimate comparison.
The new cards will offer improved tessellation performance (which we'll touch on below) as well as a new anti-aliasing feature, an update UVD section and even support for newer display configurations.
Here you can see a full Barts GPU that should look pretty familiar to those you that studied the HD 5000 series of cards and the Cypress GPU. The architecture is basically the same but has been reconfigured to improve performance per square mm of die space. You'll notice that there are now 14 SIMD engines totaling 1120 stream processors. Yes, that is fewer stream processors than the 1600 found in the HD 5870 and even the 1440 found in the Radeon HD 5850 but fear not - thanks to increased clock speeds and other changes the Barts GPU will be quite competitive.
The memory interface is still a 256-bit connection but it has been slightly degraded compared to the Cypress controller, again in order to save die space. The new memory controller has about 20% less memory bandwidth but takes up less than 50% of the die area; and while that might not be important for the $400 market of GPUs it can mean a lot for power consumption and profit margin on the $200 cards. Thus, AMD decided to use this iteration for the HD 6800 cards that will be found in that exact market segment.
What did NOT change is the number of render back ends (aka ROPs) on the GPU; the 32 found on the HD 5870 remain on the HD 6870 and HD 6850. This is basically a "rebalancing" of the architecture for AMD like NVIDIA did moving from the GF100 to the GF104 chip. By increasing the ratio of ROPs to SIMD units AMD was able to pull more efficiency out of the existing architecture as well as push the clock frequencies higher. This is a game of Jenga like you have never played. Thanks to the space saved by the different memory controller, AMD was able to keep the ROP count high.
The Barts GPU consists of 1.7 billion transistors, down quite a bit from the 2.15 billion used on the HD 5870 card (and HD 5850 though with SIMDs disabled) and thus a smaller die area (334 mm^2 versus 255 mm^2).
A quick note: while the diagram above appears to indicate that we have dual dispatch processors on the HD 6800 cards that did not exist on the HD 5800 cards, that is actually not the case. The application of the ultra-threaded dispatch processor is identical from the Cypress parts to Barts; only the way the diagram is built has changed.
Tessellation is one area where NVIDIA's Fermi architecture really excelled and in certain tests, like the Unigine Heaven benchmark, that is incredibly apparent. And NVIDIA is quick to tout their advantage and push any game or test that uses heavy tessellation on the media and consumers in order to push their advantage. While this is AMD 7th generation of tessellation of engines in their GPUs, it definitely needs some help keeping up with the work NVIDIA did on their products. AMD has tweaked the engine to a point where they are seeing improved performance at lower tessellation factors (1-11, as seen above) where games that are using adaptive tessellation engines are going to be running for the majority of the time.
You can see in the graph though that as the tessellation factor increases the difference between the HD 5800 series and the HD 6800 series will minimize quickly so we'll have to see how much better the real-world gaming performance gets between these two generations.
With all of these architectural tweaks AMD was able to improve the clock frequencies on the cards as well.
Radeon HD 6870 Specs
While the Radeon HD 5870 ran at 850 MHz and the HD 5850 ran at 725 MHz, the new Radeon HD 6870 will come with the full Barts GPU clocked at 900 MHz out of the gate. It requires a pair of 6-pin power connections though the load board power is rated at 151 watts - about 30 watts lower than the HD 5870 card.
Radeon HD 6850 Specs
The new HD 6850 will come with a 775 MHz clock rate, 960 stream processors (12 SIMD units), the full 32 ROPs and a 127 watt board power and a single 6-pin connector.
Looking at this quick comparison table new Radeon HD 6870 has about the same performance as the HD 5850 even though it has quite a few "lower" specifications. Lower memory bandwidth, fewer SIMD units, fewer texture units; but thanks to the click speed increase (from 775 MHz to 900 MHz) the new HD 6870 card will be putting up a fight.