workstation for CUDA
I am planning to build a work station with FOUR PCI express Fermi GTX
470 boards for CUDA, to run on amd64 stable. Four GTX 470 is the only
fixed hardware (Fermi Tesla has no advantage when single precision
suffices as in my case for molecular dynamics; this implies that the
intended software runs on GTX 470, like on Tesla). Thus, the culprit
is the clock, not the RAM size. However, no overclock, as the
computations may last weeks.
I am asking advice about the mainboard and CPUs (overwriting
intended). Also, if equipped with ECC RAM, it could also serve for
double precision using the CPUs only.
Supermicro 4021 GA-62R+F (with power supply 1400W)
Two Amd Opteron Six Core 2427
Actually the Supermicro cage+power_source drains a lot of money (at
least when imported to Europe) and I would prefer separate
cage+main-board+power-source (provided they fit together; three years
ago I assembled a Supermicro four-socket board starting from an old
Chinese cage; it required a full week engineering the cage, which was
for a smaller board)
Thanks a lot for advice.