The big feature of the GeForce 256 was hardware offloading of transform and lighting calculations. While it was fixed function it was programmable. This offloaded a pretty significant amount of work from the CPU and lightened the IO load single the math was done on geometry living in the VRAM.
The NV1 and other cards before the GF256 did accelerated drawing and geometry raster but lighting and transformations were done on the CPU in main memory then transferred to VRAM and rasterized.
Even if the GF256 had say the same texel output performance as a TNT2, the T&L running on the GPU would allow for more complex scenes at higher frame rates than a TNT2 doing T&L on the CPU. The GF256 also had better raster performance so for many games it was just a "free" performance upgrade over a TNT2.
The NV1 and other cards before the GF256 did accelerated drawing and geometry raster but lighting and transformations were done on the CPU in main memory then transferred to VRAM and rasterized.
Even if the GF256 had say the same texel output performance as a TNT2, the T&L running on the GPU would allow for more complex scenes at higher frame rates than a TNT2 doing T&L on the CPU. The GF256 also had better raster performance so for many games it was just a "free" performance upgrade over a TNT2.