Maybe switch to wPrime 1024M?
For NVIDIA Stage 1 I think GeForce 6/7 would be nice. Considering especially GF6 tends to be expensive making it only one score per GPU core would make it more affordable.
E.g.
GeForce 7 Series: 4 scores, 1 score per GPU, G70/G71/G72/G73
GeForce 6 Series: 4 Scores, 1 score per GPU, NV41, NV43, NV44, NV45
Bench wise 3D 2001SE might be fun but would probably require come kind of CPU limitation.
edit: One could also allow Quadro GPUs as those tend to be less expensive.