Benchmarking Half Life 2

ATI and Valve had a number of machines, all with the same base hardware configuration that Valve’s own benchmarks were taken on so we could do our own testing. However, because of the length of time it took to render the benchmark on some of the slower boards had we just done our own testing we wouldn’t have had many results due to time constraints. In order to get a wider variety of data points we shared our results with Lars from Tomshardware, Brandon from FiringSquad and Bob from Gamespot. Even Though we grouped together to get a number of different sets of data, much of it is still somewhat sparse and we’ll do some more complete testing once we have the full benchmark.

The boards tested were a 128MB Radeon 9800 PRO, a 256MB GeForce FX 5900 Ultra, a 128MB Radeon 9600 PRO and a 128MB GeForce FX 5600 Ultra. ATI had just purchased the 5600 for this test and they had requested a 5600 Ultra, but were handed a plain 5600 – fortunately Lars had a 5600 Ultra on him which we were able to use.

Note: The DX9 benchmark mode does not yet utilise Higher Dynamic Range rendering under the DirectX9 path. The point of the DirectX9 path was to give as much apples to apples rendering as possible for all boards, however present NVIDIA drivers do not support any floating point rendering targets and hence HDR would not operate on NVIDIA's DirectX9 boards. To ensure that that same amount of work is done on both ATI and NVIDIA's boards HDR is not enabled in the benchmark yet.

Note: The low resolution numbers are not present for the 5900 Ultra due to insufficient time, not due to any issues that we are aware of.

The Demo machines were actually to Valve's configuration, and they actually conspire against ATI here t some extent. First, its quite clear from the performances of the 9800 PRO, and the 9600 PRO to some degree, that the ATI boards are not actually wholly limited by their own performance, but by the CPU performance. Half Life 2 has lots of environmental, physical calculations going on, requiring a strong CPU so here the limitation for the 9800 PRO is not its shader performance, but rather the performance of the CPU - ATI would have preferred a 3.2GHZ P4 in the machine as their performance would have been higher. Also, the machines were equipped with flat panel monitor which were only capable of 1280x1024 and so they were unable to go to a resolution which would have put more onus on the performance of the 9800 as opposed to the CPU performance. This is the reason why the 9600 PRO had such large "bang-for-the-buck" ratio as it was calculated on a resolution of 1024x768.

The e3 techdemo test is actually the least of the CPU limited test for the 9800 PRO, as this does contain lots of different shaders. Both the 9600 PRO and 5900 Ultra show that they are very limited by their shader performance, although the 9600 PRO less so than the 5900. The next two demos are much more representative of levels from within the game, and although the 9800 PRO is more or less completely CPU bound in all tested resolutions and the 9600 PRO a little less shader bound, the 5900 is still showing that it is much more limited by its shader performance.

In general our testing mirrors the results as seen by Valve, yet these are subject to change – both IHV’s will have driver improvements in the pipeline. However, under the pure DX9 path it seems highly unlikely that driver improvements will make up that level of difference.

More results are available in this Excel file.