Radeon HD2900XT review

Texture units
AMD has done some revising at the texture unit level, providing four units total with eight texture address processors, 20 FP32 texture samplers and four floating point texture filter units each, supporting trilinear and anisotropic filtering for all texture formats, of which it can support up to 8192 x 8192 pixels in size.

As a result of this power, AMD claims the HD2900XT can perform about seven times faster than the Radeon X1950XT when it comes to the filtering of textures. This should open the door for higher-level anisotropic filtering. This is a bonus, as the optional high-quality AF mode available in the X1k series is now the default and only option.

All aboard the memory bus
The HD2K series uses a ‘ring bus’ architecture, but this time has gone the whole hog over the hybrid crossbar/ring bus model seen in the X1K series. The internal ring bus has a thousand wires grouped essentially into four 256-bit buses, two going one way allocated to read, and two going the other way for write. This gives an effective full duplex bandwidth of 512-bits internally, and with eight 64-bit channels to local system memory this bandwidth is available externally as well.

A bus ring stop. This is where all the read/write action happens.

Just like the previous architecture, the design allows requests to be sent around the ring bus to the nearest ring stop – if the data it’s looking for is located on the memory connected to that stop, the read/write operation is carried out. If it scores a miss, then it moves along the bus until it finds the correct stop.

The previous ring bus iteration only sent read requests through the ring, whereas the write subsystem was situated in the middle of the bus, contributing to heat. Here, interface to PCIe is simply another stop on the ring.

Aunty Alias
As seems to be the case these days, new antialiasing modes have come along with the new architecture. Custom Filter Anti-Aliasing, or CFAA, allows the card to take custom samples outside of a pixel’s boundaries. As the name suggests, it also allows customised antialiasing modes, available from the ATI Catalyst Control Center. At the moment there are only two filters available to choose from, Wide Tent and Narrow Tent, and both tend to make scenes a little too blurry.

The idea is to be an extension to NVIDIA’s unfortunately named Quincunx AA, introduced with the GeForce 3 and something that never really took off. AMD though has the advantage of being able to add potentially better custom filters as it goes – which is a good thing, as the ultimate result of AMD’s current custom AA filters is that everything tends to look a little blurry.

ATI tells us that game developers should be able to simulate the filters with DX 10 and beyond, but naturally due to hardware support using its solution should be faster. Given that devs are generally loathe to lock themselves into one bit of hardware, we can’t see the support massively lining up behind this one.

CFAA allows the GPU to take programmable samples outside of the pixel border for greater AA flexibility. At the moment, it just makes things blurry.

Tessellating features, Batman!
One thing AMD has drawn attention to is the tessellating abilities of its DirectX 10 family – which at first sounds like the Geometry Shader (GS) in another set of clothes. Or rather for those ancients who remember TruForm,
a Phoenix-styled rise from the ashes.

Initially designed as a way to offer a GS-type solution to the Xbox 360 in lieu of DirectX 10 support, tessellation is a simple concept – a triangle can be sent to the card, then subdivided into two triangles, then those triangles can be subdivided further and further, a displacement map can be applied to exact height levels from the mesh and the net effect is a significantly more detailed model derived from a simple one. In a nutshell this greatly reduces the amount of vertices the CPU needs to send the GPU, with extra vertex generation done on the GPU itself, reducing bandwidth issues and allowing huge reductions in rendering time.

The tessellating unit in the HD2900XT hints at a brilliant future for real-time 3D, assuming it makes its way into future revisions of DirectX.

Where the tessellating unit differentiates itself from the GS is it’s massively more powerful and better suited to subdivision of surfaces. The GS on the other hand is best applied to things such as deformation of surfaces and creation of incidental vertices and lines like adding hair to
a surface.

This leads to the introduction of higher order surfaces that have been available in 3D modelling packages such as 3dsmax and Maya for a long time – Bezier patches, B-splines/NURBS/NUBS, N-Patches as well as Loop and Catmull-Clark subdivided surfaces – and it seems, whatever else developers may throw at it. In all, these are just fancy terms for different mathematical methods of applying a single 2D curve to a 3D surface to raise complexity via tessellation and achieve a new level of smoothness using only a basic dataset. In English, it means we can take a small amount of input data like a pyramid, and rather than having a straight line between the points we can make it curve however we like, with extra detail.

This also opens up the floor for dynamic tessellation, depending on point of view – meaning that low detail and high detail models are exactly the same, cutting down on 3D artist time. Different degrees of tessellation and hence detail can be applied depending on how close the player is to that object, achieving the same effect. Further down the track this could be potentially used as a performance boost for low-end graphics cards, dynamically adjusting the level of detail down to maintain fps where necessary.

Early Windows Graphics Foundation (the initial name for DirectX 10) documents reveal tessellation as an optional feature of the spec, and all suggestions lead towards it becoming a mandatory part of the next revision – we certainly hope so, as the advantages are huge. For now, AMD is making the tessellator available through calls to the vertex shader.