A new GPU for a new DirectX
Let's be perfectly honest here: even though the ecosystem of graphics cards has been getting cheaper for consumers, and we have all been able to get mounds of gaming performance for a very low price, the last time we have seen a full and complete new architecture was with the AMD RV770 architecture back in June of 2008. Before that was the NVIDIA GT200 architecture that preceded it by only a couple of weeks. The two designs couldn't have been more different: NVIDIA went with a huge die size that destroyed AMD for the enthusiast-class single GPU performance while AMD went with the much more power efficient model of creating a GPU for performance segment and scaling both up and down as price points permitted. While the debate would probably continue for weeks if we decided to retroactively declare a winner in that fight, but my opinion sides with AMD overall execution - they created great parts, lowered prices and took a look of market share from NVIDIA's once tightly gripped fingers.
Now, 15 months later, AMD is the first company back with a new design primed and ready for the DirectX 11 API while adding performance in significant amounts. Hardware enthusiasts and news hounds have been hearing about RV870 and/or the Evergreen class of GPUs for quite some time but the team at AMD did a surprisingly good job keeping much of the detail close to their chest until today.
In my review of the ATI Radeon HD 5870 graphics card we will dive into the DX11 feature set included on the new architecture, explain how this design has changed from the RV770-based core, talk about AMD's new Eyefinity multi-monitor gaming technology and of course dive into the benchmarks and power consumption numbers that will help us to actually decided is all the work AMD has put into the design was worth it.
A new card for DirectX 11
When we first saw the Evergreen architecture demoed at Computex this past June, AMD was already riding high on the claim that its new RV870 architecture would be the first to support the new DirectX 11 API from Microsoft and that NVIDIA would be months behind them in execution. Even if that is the case, AMD needs to prove to gamers and developers that DX11 is something WORTH having, and hopefully worth having before NVIDIA's responses are available.
AMD is proud of its architectural advancements and "firsts" as they like to call them including the introduction of GDDR3, GDDR4 and GDDR5 memory technologies as well as new process technologies like 65nm, 55nm and the now infamous 40nm from TSMC. Technological firsts are great but only if they result in a better product than the competition; that didn't occur in every instance though.
We also know very well that computing workloads ARE changing and the advent of software technologies like OpenCL and DX Compute are going to be able to move more and more processing from solely running on the CPU to running on both the CPU and GPU more efficiently. AMD continues to be in a great spot for this change, owning both a solid CPU solution in their Athlon and Phenom lines as well as the entire Radeon team.
Along with those changes in computing methods comes Microsoft's attempt to standardize GPU computing called DirectX Computer. The new Radeon HD 500 series of parts will be the first to support new Shader Model 5 features and is already well on its way to becoming OpenCL 1.0 and possibly even 1.1 compliant. And as this slide points out, in terms of raw compute power, the new Evergreen GPUs will have nearly triple the single precision floating point power compared to NVIDIA's current GT200 parts.
Where will that power go? AMD is telling us, as are the individual developers, that DX11-based content is a lot closer than you think including titles that implement DX Compute! You can see that some key titles like the new "Aliens vs Predator" and upcoming "Stalker" title will at least by using SOME kind of DirectCompute features.
There are other features that are being introduced with DirectX 11 that are interesting including a required tessellation engine, better texture compression and new multi-threaded performance advancements. I know that many users will point to the lack of development for DX10 as a reason to simply shrug off DX11 but that likely won't be the case this time around. The runtime for DX11 fully supports DX10 hardware and the new API will NOT be exclusive to Windows 7 - it will be ported to Vista sometime this year.
The new Shader Model 5.0 support includes new instructions that improve sample coverage information for pixel shaders, data fetches of sample values in single textures and more that will benefit game developers with applications for edge detection AA, faster shadow filtering and occlusion, etc. There are TON of details on these new functions and their uses if you are interested a simple look around Google will net you more reading that you probably want to take on.
We mentioned a few games that were implementing DirectCompute 11 above and this list here tells us some of the areas those titles and other upcoming games will optimize for. Obviously the sexiest answers are for physics, AI and ray tracing, though information isn't really abundant yet on any implementations of these potential advantages. For example, it's likely that middle-ware companies will develop a DX Compute physics engine that will compete with NVIDIA's PhysX - that would great a great open-standard for GPU-accelerated physics going forward.
Even though tessellation has been a feature in AMD's GPUs since the days of the HD 3800, it has yet to be used in any meaningful way. During our time with developers last week we heard from at least a couple that were excited about the benefits of a DirectX-based tessellation engine standard.
Tessellation allows developers to create additional geometry data with very little performance hit using dedicated hardware for the functionality. In the image above you can see the base level wireframe of the Alien and Predator and then a tessellation-enhanced version on the right. Clearly there is more detail in the Alien model that will smooth out the look of the character as the camera zooms in.
Order independent transparency allows for GPU-based calculations on the z-depth of different characters or sections of characters and thus allows for stacked transparent objects to be rendered more realistically than simple alpha blending.
Obviously the benefits of render post-processing are laid out here pretty plainly - effects like depth of field, motion blurs and smoothing could all benefit from this API advancement. Think of it as the ability to run a Photoshop-style filter across the entire screen for each frame.
Ambient occlusion is a very computer-intensive task and render post processing provides another avenue for that image quality enhancement as well as creating shadows that are dynamic based on their physical relation to the subject.
As I mentioned before, AMD wants everyone to believe that DX11 will be relevant sooner rather than later so that their advantage with an earlier release of DX11-ready hardware is not wasted. They offer up this exhibit: games coming out in the near future that utilize DX11 in some fashion. Obviously the killer titles here are DiRT 2, STALKER and Aliens vs Predator, but how much of an advantage you get from DX11 on each of these titles has yet to be seen.
Luckily for AMD, the new Evergreen series of GPUs has more to stand on that just DirectX 11.