The card (quadro K2000D) is -unfortunately- not up to required standards, but it was possibly the only reasonable solution : not only the space in the riser PCI assembly of the server is very tight (requiring a slim card), but the 16x slot can only take up to 70 Watts worth of a card. This means GTX 645 (at ~120W) wouldn't do it, and K4000 (rated at 80W) was too expensive at ~$800.

Anyway, here are the numbers :

6,500 atoms

04 cores : without CUDA → 15.2 ns/day , with CUDA → 12 ns/day

25,000 atoms :

04 cores : without CUDA → 3.64 ns/day , with CUDA → 6.90 ns/day

08 cores : without CUDA → 6.96 ns/day , with CUDA → 7.26 ns/day

16 cores : without CUDA → 11.62 ns/day , with CUDA → 7.27 ns/day

60,000 atoms :

04 cores : without CUDA → 1.29 ns/day , with CUDA → 2.70 ns/day

08 cores : without CUDA → 2.43 ns/day , with CUDA → 2.97 ns/day

16 cores : without CUDA → 4.50 ns/day , with CUDA → 3.10 ns/day

The way it looks, it would possibly make more sense to place the card in an old quad-core box (n0006 ?) where it would come handy for a simulation with a large number of atoms, instead of wasting it on the IBM server (why running a job with only 4 or 8 cores in a 48-core machine ?).

⇒ Take-home message : stick to simple boxes (with ~16 cores) with very fat nvidia cards. For current versions of NAMD they can produce as many ns/day as a 48-core machine without a fat nvidia card.