Gaussian 90 SCF times

[m10!frisch (- at -) uunet.UU.NET (Michael Frisch)]

 On the subject of workstations for ab initio calculations, one person
 commented:
     We have been having very good experience with the Silicon Graphics
     RISC servers, running both Gaussian-90 and GAMESS (as well as a
     host of other numerical simulation, scattering theory, and other
     heavy computations).  Running on one (of the 4) CPU of a year-old
     4D/340 with 33Mhz CPU's, a large (ca. 850 Mb of scratch) SCF
     (G-90) calculation which took 100 minutes on a CRAY-XMP, ran in
     only 300 minutes on our SGI box.  I understand that < $20K 1-CPU
     versions of this machine WITH graphics are now available, and from
     our experience, they would also be very robust multi-user servers;
     we are usually running with several background jobs, plus several
     interactive users using both dumb terminals and X-terminals, and
     notice no real problems.  These machines can also be readily
     goosed up with cheap memory (e.g., $982.Cdn/8Mb from Kingston) and
     SCSI disks (ca. $3300.Cdn/1.2Gb) from 3-rd parties.
 The SGI is a good machine, however, I'd like to point out that an SCF
 job which takes 100 minutes on a Cray in Gaussian 90 using
 conventional SCF typically takes half that much CPU time when run in
 direct mode.  That doesn't change the above conclusions, but we would
 like people to get the most out of their Crays as well!
 Michael Frisch
 Gaussian, Inc.
 -------