If you make your demo to work on one ATI card you will have a small audience at most, and the card will be out of the market in a year and all the tricks you learnt were for nothing. And if you want it to work on more hardware you get into the device driver business, and that rules out special tricks for specific hardware to do non-standard rendering anyway...
Back in the days of home computers there was one longer-lived hardware platform that many people were using. There is a faint hope that initiatives such as the Raspberry Pi may bring people to "standardize" on one more or less ubiquitous platform for longer time, but I'm afraid this ship has sailed. So it makes sense for demosceners to focus on what is possible on the higher, portable API level.
You program the GPU by writing shaders that run right on the GPU, the only perf increases to be had involve how fast you can move data on and off the GPU. Eliminating all abstraction might give you a few percentage points, not enough to do anything that we would consider amazing under current circumstances.