Yup, the original work was done by them, but the really impressive bit is still down to ATi if they managed to take that simulation, convert it to GPU runnable code, and render it real-time at a decent resolution..
The original video took a bit longer:
We were able to simulate computational grids with effective resolutions as large as 256 × 256 × 192 for the fluid and as many as 90k triangles for the rigid and deforming bodies using a 3 GHz Pentium 4. The computational cost ranged from 5 to 20 minutes per frame, and thus the longest examples took a couple of days.
Like I said, if that's running in real time at a decent resolution, then ATi deserves a *lot* of credit.