>You may be right - I was reacting from earlier experience
>porting Amber in single-precision. Which gives me a thought:
>if the ES-40 is somehow running single precision, those subtle
>deviations in all the numbers would be exactly what one would
>expect. You should be seeing "-DDPREC" in the compilation
>statements; if not, check src/sander/Makefile, where double
>precision is enforced by "PREC=DPREC"
Accuracy on 64 bit platforms shouldn't be different from 32 bit
platforms in double precision mode because the "double" is 64 bits.
If the ES-40 is a 64-bit machine, then my remarks are even
less founded than I thought.. unless the machine has a
single precision emulation mode that could get turned on
accidentally (unlikely), I would begin to suspect a hardware
problem. Have you tried repeating the same run to see if
anything changes? Can you force it from one cpu to another
in single-cpu mode? Are there other diagnostics that could be
run? How about compiling w/ no optimization?
Bill Ross
Received on Sun Jun 24 2001 - 23:51:20 PDT
