At least for my NVIDIA GPU, it has bootup clocks in BIOS, the clocks at which the card boots and operates, until there's a working driver (Windows/Linux/Mac) which then sets the normal clocks accordingly to the BIOS, those two sets of clocks are entirely separate. Those bootup clocks are set to very low, as to not overheat the GPU immediately. You can see the bootup clocks, if you use Safe Mode/LiveCD Windows environment, and try to obtain the current GPU speeds.
Since bootup clocks are not meant to be changed (only by using a BIOS Editor), you get the same performance, since as soon as the Windows driver is shut down, the clocks revert to bootup clocks, and in DOS, you will always get the same performance, no matter what you set in the PowerStrip.
This is how it works on my 6800GT. I do not know if this applies to older NVIDIA cards/Ati/3dfx/whatever else.

DOS performance depends on basic CPU to bus to card transfer... DOS programs don't care about the card's windows acceleration functions, and it's those that enable the card to perform at the speed of its GPU & RAM, rather than the speed of the bus.

Matth79 wrote:

DOS performance depends on basic CPU to bus to card transfer... DOS programs don't care about the card's windows acceleration functions, and it's those that enable the card to perform at the speed of its GPU & RAM, rather than the speed of the bus.

So what determines the performance differences that we do see between cards in DOS benchmarks? Like, what is it about, say, a VLB S3 805 that makes it so much faster than a Weitek P9000 in DOS?