Missed it, big time . You may want to read up on past posts v.v. 32 versus 64 bits, for CEP2 too!.. the overall gain is small [4-8%] and can turn out even negative , yes ! Why 64bit exist for a substantial part is systems that can´t do 32 bit apps.

Whilst, it's a waste discussing CEP2 in it's current incarnation as a new version is in development, using a faster/advanced Q-Chem incarnation, that will also bring much bigger models to the grid... prep for 4 to 8GB. No, this is not unique for the grid... QMC [Quantum Monte Carlo] has near 5GB models running. If you don't have the hardware for the big ones, don't tick the opt-in box. Running one of those will be a challenge, running them on multiple cores... try not to use the system for anything else requiring display feedback. If that comes in 64 bits... we'll learn when the alpha testing is ready and beta is started on the grid.

Nothing new I posted, but like to add, before the tilt function kicks in faster than hyperwarp, these big models are a subset of what is going to run on CEP2 v2, so to speak. The standard model sizes remain and [refrain from expressing any confidence at all it will be], another opt-in button in the project selection section. Since apps test RAM before sending, not probable they'll be going to be release to fringe device, who've set big VMs. The IO will be...fill in the blanks... and since there's only 16 checkpoints ** when the intermediate result is smallest, not expecting much more writing in-between... something to profile when we get to Beta test them.

Side note, the version will then get the major of 7.xx where we are now on 6.40, where the 6 is referring to the BOINC API used for the wrapper build. 7.xx thus fully suited for BOINC v7 [held me stethoscope with amplifier to the binaries, but not a sound of when :P]

edit: ** That was an poor assumption. Of course we [me] does not know if the whole checkpointing thing is going to be different [within the range of possibles naturally]. Something else we learn when BETA is there, that is, if we get full models to BETA and not anything sliced down to speed testing up.

My point is that 64-bit apps can access more than the first 4GB of memory. Assuming the quoted memory requirement (yes, I know it's worst case), running more than two or three instances will cause any more to rely heavily on a Windows swapfile. Having a 64-bit version of the app would allow additional instances to run primarily in memory, with minimal paging. You'd have to work very hard to convince me that using addtional main memory is slower than disk access.

Correct me if I'm wrong, but since each work unit runs its own process it doesn't matter how many 1GB 32 bit apps your running as long as your system has enough RAM