mccalpin@masig1.ocean.fsu.edu (John D. McCalpin) (07/31/89)
I have been concerned about the floating-point performance of the NeXT, so now that I've got one to play with, I decided to see how bad it really is. The FORTRAN results below were compiled on a Sun-3 with the -O and -f68881 options. The resulting executables were copied to the NeXT and run through 'atom (1)' to make them executable under Mach. The C results below were compiled on the NeXT with the standard (GNU) compiler. The Sun results used Sun's (ugh) compiler. LINPACK performance in kFLOPS NeXT Sun 3/260 w/68882 w/68881 w/Weitek 32-bit Fortran LINPACK : 227 130 860 64-bit Fortran LINPACK : 203 110 460 32-bit C LINPACK : 240 104 446 64-bit C LINPACK : 210 98 249 I also have a "favorite" benchmark code from one of my applications. A small excerpt of my results shows for 32-bit Fortran (except the Cray): seconds ratio to MicroVAX NeXT 158.4 1.23 Sun 3/280 (fpa) 55.3 3.53 Sun 3/280 (68881) 275.2 0.71 IRIS 3130 (fpa) 151.4 1.29 MicroVAX II 195.4 1.00 DECstation 3100 15.5 12.61 Cray X/MP 1.9 102.84 So for a cheap machine with a 68882 coprocessor, the NeXT does quite well - especially in C. Of course, if you really want to crunch, the DECstation 3100 has MUCH better floating-point performance for only a few thousand more $$$. -- John D. McCalpin - mccalpin@masig1.ocean.fsu.edu - mccalpin@nu.cs.fsu.edu mccalpin@delocn.udel.edu
blenko-tom@CS.YALE.EDU (Tom Blenko) (07/31/89)
In article <MCCALPIN.89Jul30144334@masig1.ocean.fsu.edu> mccalpin@masig1.ocean.fsu.edu (John D. McCalpin) writes: |I have been concerned about the floating-point performance of the ... |The FORTRAN results below were compiled on a Sun-3 with the -O and |-f68881 options. The resulting executables were copied to the NeXT |and run through 'atom (1)' to make them executable under Mach. | |The C results below were compiled on the NeXT with the standard (GNU) |compiler. The Sun results used Sun's (ugh) compiler. So are you comparing compilers with compilers? | |I also have a "favorite" benchmark code from one of my applications. |A small excerpt of my results shows for 32-bit Fortran (except the Cray): | seconds ratio to MicroVAX | NeXT 158.4 1.23 | Sun 3/280 (fpa) 55.3 3.53 | Sun 3/280 (68881) 275.2 0.71 Sun+GnuCC would be more interesting. And how do disk speeds compare? |So for a cheap machine with a 68882 coprocessor, the NeXT does quite |well - especially in C. Of course, if you really want to crunch, the |DECstation 3100 has MUCH better floating-point performance for only a |few thousand more $$$. If you are a university making a major purchase (75 nodes or more, so I hear), you should be able to get a DECStation or a SparcStation with 8Mb, B&W, 100M local disk for under $5K (I know of three universities that have done just this). Which means you can get much better FP performance for significantly less money. Yet another reason why I say a NeXT should cost about $4K. Tom
mccalpin@masig1.ocean.fsu.edu (John D. McCalpin) (07/31/89)
In the first referenced article above, I posted some floating-poing
performance results for the NeXT.
In the second article referenced above, Tom Blenko (blenko-tom@cs.yale.edu)
made some replies.
me>|The C results below were compiled on the NeXT with the standard (GNU)
me>|compiler. The Sun results used Sun's (ugh) compiler.
tom>So are you comparing compilers with compilers?
Most of my results are comparing hardware with hardware, since the
same f77 compiler was used. I will post the equivalent numbers for the
Absoft f77 compiler when they become available.
I used the default C compiler's on each machine. If Sun chooses to ship
a crummy compiler, that is not my problem. I do expect Gnu C to do better
on the Sun, but I do not think that it is installed on any of the machines
that I have access to.
me>|I also have a "favorite" benchmark code from one of my applications.
me>|A small excerpt of my results shows for 32-bit Fortran (except the Cray):
me>| seconds ratio to MicroVAX
me>| NeXT 158.4 1.23
me>| Sun 3/280 (fpa) 55.3 3.53
me>| Sun 3/280 (68881) 275.2 0.71
tom>Sun+GnuCC would be more interesting. And how do disk speeds compare?
(1) Sun+GnuCC would not have told me anything about the f77 performance
of the two machines. The overwhelming majority of number-crunching
codes are still written in FORTRAN.
(2) The disk speeds are not a noticeable factor in the above comparison.
Less than 10% of the time is spent doing I/O, and most of that is
spent formatting the output.
me>|So for a cheap machine with a 68882 coprocessor, the NeXT does quite
me>|well - especially in C. Of course, if you really want to crunch, the
me>|DECstation 3100 has MUCH better floating-point performance for only a
me>|few thousand more $$$.
tom>If you are a university making a major purchase (75 nodes or more,
tom>so I hear), you should be able to get a DECStation or a SparcStation
tom>with 8Mb, B&W, 100M local disk for under $5K (I know of three
tom>universities that have done just this). Which means you can get
tom>much better FP performance for significantly less money. Yet another
tom>reason why I say a NeXT should cost about $4K.
The heavily discounted prices that I have seen on DECstation 3100's
still place them at over $10K (with 19" monochrome monitor, 8MB RAM
and a 330 MB disk). I don't know about the discounted SparcStation
prices, but its floating-point performance is not competitive.
--
John D. McCalpin - mccalpin@masig1.ocean.fsu.edu - mccalpin@nu.cs.fsu.edu
mccalpin@delocn.udel.edu