1k1mgm@kuhub.cc.ukans.edu (Christopher Gunn) (04/12/91)
Yesterday I visited our main vendor for molecular modeling software (Sybyl, Tripos Assoc., St. Louis) to get some hands-on experience with various workstations we're evaluating for purchase. I'd done some test runs on our university's VAX 9000-210 of models I planned to run, but shrunk the iteration counts down quite a bit so I wouldn't tie up my host's little workstations for TOO long, right? Test reproduced below was a conjugate-gradient-type energy minimization of a 44-residue peptide, 40 iterations; Tripos' code is written in portable C (ANSI or close). Results: Platform Time Rel. SPECs Discrepancy Notes (clock;cpu) (sec) Speed VAX 9210 554 1.0 ~55* ---- ESV50 616 0.9 ~14** ~3.5 Machine busy*** (25Mhz R3000) SGI 4D/210 434 1.3 14 ~5 Machine idle**** (25Mhz R3000) IBM RS/6000 269 2.1 29 ~4 Machine idle 530 (25Mhz) *SPEC value from memory **SPEC value is guess--I don't have numbers ***ESV heavily loaded--maybe 25% of CPU available for job ****CPU radically idle, since this was really a 240 and the job didn't even have to share processor with OS I've got some other numbers I'm still trying to digest; may post them if there's any interest. These results messed with my head. I knew that Sybyl C code didn't fit confortably in VAX architecture--I was getting only 39 microVUPS from the 9210 vs. 60-80 microVUPS for MOPAC, Amber, etc., runs on 9210, but finding that these various RISCS were roughly FOUR TIMES faster than I thought for my application was a severe (but obviously pleasant) surprise. Does this make sense? Is there anybody out there with number- crunching code in semi-elegant C that flies on RISC, dies on VAX? Also: any experience pro/con on ESV3 or ESV5x? Graphics on these things scream, even though (for Sybyl) code is in PEX, vs. native GL for SGI and IBM. Christopher Gunn Molecular Graphics and Modeling Lab SPAN--KUPHSX::GUNN Department of Medicinal Chemistry, Malott Hall 913-864-4428 or -4495 University of Kansas, Lawrence, KS 66045