guppy@orville.mit.edu (Harold Youngren) (02/21/91)
I recently bought one of the new 400Mb NeXT slabs from the MIT micro-computer center. I am involved in CFD (computational fluids) code development and use, principally on DEC 5000 and Stardents here at MIT. These are mostly big FORTRAN codes. Postings on this newsgroup have indicated that the NeXT was quite fast in comparison to other workstations. Unfortunately the information on Fortran has been kind of sparse, although the f2c translator package has been mentioned several times. Here are some speed comparisons for f2c/cc code on several machines. These results were obtained using two flow codes, a small unsteady flow code that solves a 100x100 complex matrix, and a huge viscous turbomachinery code that solves 5K x5K banded matrices. The results were essentially the same for both. Machine Speed Claimed MFLOPS Single/Double NeXT 1.0 ?/2.0 DEC3100 1.9 4.0/1.6 DEC5000 3.3 6.4/3.7 VAX VS2000 0.22 ? The speeds were normalized by NeXT speed, obtained from user+sys timings with >time xxxx. The f2c code was the latest version from research.att.com and was compiled for use on both MIPS and NeXT. The NeXT had 8Mb of memory but paging did not seem to be a problem on the big code. Timings comparing f2c/cc with native f77 compilers show that the translated code is roughly 20%-30% slower that a full f77 compiler. I am interested in hearing about experience with the Absoft compiler, which is not cheap ($750 academic), but may be the only other game in town. On a pure price/performance basis the DEC3100 or DEC5000 on an academic discount outperforms the NeXT AND comes with good native compilers and general acceptance of networking standards, ie. X built-in (also color). The user unfriendliness of X and DEC's UNIX is another issue, however:-). Harold Youngren guppy@orville.mit.edu