Radeon HD2900XT review

High definition tech
The HD2900XT is a long card at 241mm – this puts it squarely in between the 8800 GTS and GTX. It’s quite heavy as a result of its heatsink, and very securely put together with a metal bracket on the back, giving it a weight of 952g compared to the 8800 GTX Ultra’s 766g and the 8800 GTS’ 684g.

It follows the same ‘blower’ configuration on the heatsink as NVIDIA’s parts, with a 75mm fan pushing hot air out of the case as well as from the GPU. Curiously a second header is featured on the board – potentially to power more exotic cooling solutions like a water-cooling pump as well as a fan.

Two power connectors are present – one eight-pin and a traditional six-pin. If you’re in doubt as to whether the industry will move to eight-pin power or not, PSUs are already on the shelves with dual eight-pin connectors, and the 8800 GTX Ultra has the empty holes ready in the PCB to support an eight/six-pin configuration like the HD2900XT, despite being a dual six-pin solution.

Unlike NVIDIA’s odd 320-bit/640MB for the GTS and 384-bit/768MB for the GTX, ATI have opted for a 512-bit memory bus and 512MB of GDDR3 – which should get interesting once we get to high resolution, large texture games.

Caps to pop
There is a grand total of three capacitors on the board, greatly reducing the risk of them being torn off or causing problems.

This risk is reduced further by only two of these capacitors being exposed outside the heatsink.

The secret behind this is digital Pulse Width Modulation, taking on the responsibility of power stabilisation to the core and memory and hence minimising the number of capacitors. This also has the run-on benefits of resistance to higher temperatures, higher efficiency and an easier to work with PCB.

Architecture

The R600 is quite the engineering feat.

Of course just like all good DirectX 10 parts should, the HD2900XT supports Shader Model 4.0, complete with a unified shader architecture. ATI has had experience with this before in the Xbox 360’s Xenos GPU – and so unsurprisingly this heritage comes through with the R600.

The Ultra-Threaded Dispatch Processor can indeed handle many threads.

Much like NVIDIA’s 8800 series, a thread processor is employed to manage exactly how the unified shaders and texture units are used. In this case AMD calls its solution the ‘Ultra-Threaded Dispatch Processor’. This feeds a thread – consisting of vertices, polygons or pixels – through to the 320 stream processors present on the HD2900XT, arranged as four SIMD blocks containing 80 processors each. Two arbiters and sequencers are allocated to each SIMD block, reducing latency by being able to feed texture and vertex fetch commands into each, while preparing to push data into available stream processors. If a vertex or texture thread stalls, then whatever data the other arbiter holds can be pushed through instead, rather than having to go all the way back to memory. As such, latency is greatly reduced simply by being able to process a number of threads at once.

The stream processor block. There are 64 of these in the system containing five scalar units, leading to 320 stream processors in total. The thicker floating point unit on the left can handle extra operations like SIN, COS, LOG and EXP on top of the usual multiply/add instructions.

When finished the work hits the render back end units – or ROPs, as they’re more commonly known – and eventually makes its way to screen.

At this point more astute readers will have picked up the 320 stream processors compared to NVIDIA’s 96 on the GeForce 8800GTS. There are a few differences to point out here.

The texture units are more texture processors now, given what they’re capable of.

Firstly, NVIDIA’s stream processors run on a separate clock. In the GTS this is 1.2GHz. AMD’s solution on the other hand is tied to its core speed, meaning on the HD2900XT they’re set to 742MHz – about 1.8 times slower. Still, it’s not as simple as that. NVIDIA says it counts just standard ALUs, while AMD includes both standard and special function ALUs. By this measurement NVIDIA claims it would have double the shader units, at a significantly higher clock speed.

The render back end, or Render Output unit (ROP) of the R600. The addition of a new programmable multi-sample AA block is what allows CFAA.

AMD on the other hand argues that as the whole card runs faster than the NVIDIA counterpart, overall communications should be quicker, that its architecture is lower latency, handles long shader instructions better – in fact only limited by memory size – and, of course, emphasises general purpose GPU (GPGPU).

So as it stands with two radically different approaches, we’ll simply have to let the benchmarks speak for themselves.