Earlier, I predicted that the performance of the graphics card would improve but at the expense of the processor's. The results seem to agree. Let's go through them again :-

WinBench 99

The processor tests of this benchmark were useless for the purpose of evaluating the effect of the loss of L2 cache bandwidth to the video RAM. Because they made no use of the graphics card at all, the L2 cache was not needed to cache the A0000h-AFFFFh region and thus, there was no loss of L2 cache bandwidth.

The graphics tests were more useful. As expected, enabling the feature gained a boost in the Business Graphics WinBench test but it was really surprising to see that it also caused a drop in the High-End Graphics WinBench test. That tells us that (at least with business applications), the feature is a double-edged sword.

Winstone 99

In this real world test of business applications, enabling the caching feature caused a drop of 0.5%. While the percentage isn't big at all, remember that in such a test, the graphics component is merely a part of the entire equation. This often conceals the change in performance that actually occurs. For that, we rely on dedicated tests for the performance of the graphics card, like WinBench 99 and 3DMark 99 Max.

3D WinBench 99

In the overall score, the caching of video RAM boosted performance by 0.2%. At least in this synthetic 3D benchmark, Video RAM Cacheable appears to do some good. In the various detailed tests, it was the enabled setting that won most of the top spots.

As with WinBench 99, the processor tests of 3D Transformation; and 3D Lighting and Transformation were useless in determining the effect of the loss of L2 cache bandwidth because the graphics card was also not used in those tests.

3DMark 99 Max

In the tests of all three different resolutions, enabling Video RAM Cacheable incurred a drop in overall 3D performance. This is more pronounced at the resolution of 640x480 and decreased as the resolution increased.

The processor test here is suspected to be affected by the loss of L2 cache bandwidth at the very beginning of the test since there's bound to be remnants of the data from the video RAM in the L2 cache, left over from the Game 2 benchmark that ran just before the processor test. In all cases, enabling Video RAM Cacheable reduced the processor's performance a little. Note that the results are not an accurate indicator of the degree to which the loss of the L2 cache affects the processor since the test itself does not use the graphics card at all.

Meanwhile, enabling the feature was shown to improve the performance of the rasterizer (as expected) and the texture rendering speed of the card. Unfortunately, the boost in the performance of the rasterizer was always much smaller than the loss in processor speed. Also, the gain in texture rendering speed dropped as texture size increased. In fact, the disabled setting equaled or surpassed the enabled setting beyond 8MB (640x480 and 800x600) and 16MB (1024x768).

In the two simulations of real-world games (one single-texturing and the other dual-texturing), the disabled setting, except in one occasion, was faster than the enabled setting in all three resolutions.

MS Direct3D Test

In this simple benchmark, three factors were tested - the card's fillrate, the polygon throughput from the processor and the intersection throughput. The fillrates were equal for both settings while the enabled setting was slightly faster in the intersection throughput test.

Polygon throughput, however, showed the extent of the loss in processor performance due to the diversion of L2 cache for data from the A0000h-AFFFFh region of the video RAM. This is because the graphics card was used in this test and so, the L2 cache was actively caching data from the A0000h-AFFFFh region of its video RAM. A drop of 1.1% in polygon throughput was registered when the feature was enabled, making it the most significant change recorded so far.

Expendable

Overall, both settings scored rather similarly though the disabled setting was slightly ahead. Disabling the feature allowed for a less variable (more stable) framerate that also has both a higher minimum and average framerate as well, allowing for less jerkiness. However, as the resolution increased, the performance advantage of disabling this feature decreased.

Turok : Dinosaur Hunter

In this single-texturing game, enabling the caching of video RAM decreased the average framerate by about 0.5%. In all three test runs, the disabled setting also scored both the highest maximum and minimum scores. So, the disabled setting is clearly the winner here.