I almost wrote FP64 but since there's no dedicated FP64 unit in most GPUs except things like the Quadro A series (and the AMD equivalent), I didn't.
AFAIR Fermi had quite good FP64 support by linking to FP32 ALUs. With Kepler FP64 went back to the SFUs and the Geforce driver even lowered the FP64 throughput on purpose, so the scientific community had to buy expensive Tesla and Quadro cards with the same GPUs, to get good compute performance for simulations etc.
With Fermi a cheap Geforce 480 could compete with Quadro 6000 on CUDA compute power (but did not have enough VRAM). The Geforce 580 with 3GB was really nice. Used them for raytracing my hobby stuff at home. Still have two of them lying around.
The Quadro7000 with the same GF110 GPU was only available in a QuadroPlex setup. AFAIR Nvidia never ever announced it as a standalone product.
Now even this stuff feels retro. I still remember the excitement when getting a G80 card and did my first GPGPU stuff with CUDA 0.2 alpha. Every little mistake crashed the machine 
Well even NV40 and shader model 3 now feels like it came shortly after I left the Amiga.
