tj@xn.ll.mit.edu (Thomas E. Jones) (01/05/91)
I and a friend were wondering about this new NeXTstation. The sales brocures look great. They say it's capable of processing at 15 MIPS and 2 MFLOPS. That sounds like it should be a peak speed (salespeople talking there.) Has anyone done any real benchmarking? How does this compare to a sparkstation or a Sun 3? MIT people can get an 8 MEG NeXTstation with 105MB and software for $3450. This isn't terribly different than what we can get for a sparkstation, which is is said to have 15.8 Dhrystone MIPS, and 1.7 MFLOPS double-precision LINPACK. I get the idea Sun is leveling with us, not just telling us maximum theoretical speed we can never acheive, but the real Benchmarks (they also say 10.0 SPECmarks, which I've never heard of.) If anyone has really used a Sun3 or a Sparkstation, I'd really like to hear from you. I figure the people who shelled out big money for a NeXT station wouldn't want to admit it's slow, so someone who has alternated between the NeXT and a RISC machine of some sort would be more unbiased. - Thomas E. Jones tj@xn.ll.mit.edu -- tj@xn.ll.mit.edu or tj@ll-xn.arpa (one of these should work) Thomas E. Jones, home (617) 924-8326 work (617) 981-5093
riordanmr@clvax1.cl.msu.edu (Mark Riordan) (01/05/91)
In article <1991Jan4.182011.28485@xn.ll.mit.edu>, tj@xn.ll.mit.edu (Thomas E. Jones) says: >Has anyone done any real benchmarking? How does this compare to a sparkstation >or a Sun 3? I haven't done any *real* benchmarking yet, but I've been planning on doing so. However, since you ask, here are the results of some not-so-real benchmarks: All benchmarks are number of seconds of user time as reported by "time". pi.c -- program to compute Pi using probablistic techniques. Lots of floating point multiplies and divides, some integer, some subroutine calling. SPARCstation 1: 44.6 seconds NeXTstation: 20.4 seconds Convex C240: 7.5 seconds chkprime.c -- program to verify that a given number is prime. Run on a 342-bit prime number. Lots of integer arithmetic, especially multiplies. SPARCstation 1: 232.1 seconds NeXTstation: 114.2 seconds Convex C240: 49.0 seconds newdes.c -- program to encrypt a file and uuencode it. Lots of logical operations. No floating point, no integer multiplies or divides to speak of. Fair number of function calls. Run on a 200KB file. (Note: *not* a DES implementation.) SPARCstation 1: 2.1 seconds NeXTstation: 1.0 seconds Convex C240: 2.4 seconds The Convex would have done better if I'd given it some vectorizable applications. So far, the NeXTstation has held its own in these tests, though. None of the tests is big enough to flush a cache. That's why I say the benchmarks aren't "real". Caveat emptor, etc. Further note: I'm using the standard SunOS 4.1 C compiler, not the extra-cost SPARCcompiler usually used for benchmarks. All compiles done with -O. Mark Riordan Michigan State University riordanmr@clvax1.cl.msu.edu
mahesh@caradhras.cc.nd.edu (Mahesh Subramanya) (01/05/91)
Over the last year, I've had extensive use of a SS1/SS1+, a NeXT, a DecStation 3100, a Mac IIci, a Mac IIfx, and an IBM R/600 Model 530. I've had these (in various combinations) on my desk at one point or the other for extended periods. Also, I've had a Nstation on my desk for around a week. Tho I haven't run some of the "tried and proven" bmarks on these boxes, I have done quite a significant amount of "real world" usage on these boxes. To make it short and sweet, the order of speed is ss1 < ss1+ < DecStation < Nstation < R/6000 320 < R/6000 530 Within this, the most obvious performace gap is in the SS1s. The SPARCstations seem to be the least powerful of the lot by a rather large factor. In fact, from my impression, tho' they *do* achieve their 1.4/1.7 LINPACK ratings, they usually average 1 MFLOP on most other things. By and large, SPARCS are rather slow beasties compared to their brethren. The Decstations are usually around 1.5 times faster than the SS1/1+. THe surprising result was that in most of my usage, the Nstation was faster than the DecStation!! (also corroborated by some chap in comp.benchmarks using Mathematica bmarks!). Definitely not to my surprise, the R/6000s blew the doors off the rest of the competition. -- ************************************************************************ Mahesh Subramanya INTERNET: mahesh@caradhras.cc.nd.edu Senior Analyst Office of University Computing NeXT: mahesh@numenor.next.nd.edu University of Notre Dame Voice: (219) 239-5600 x6421 Notre Dame, IN 46556 ************************************************************************
bchen@nntp-server.caltech.edu.UUCP (Bing-Qing Chen) (01/05/91)
From article <1991Jan4.182011.28485@xn.ll.mit.edu>, by tj@xn.ll.mit.edu (Thomas E. Jones): > I and a friend were wondering about this new NeXTstation. The sales brocures > look great. They say it's capable of processing at 15 MIPS and 2 MFLOPS. > That sounds like it should be a peak speed (salespeople talking there.) > > Has anyone done any real benchmarking? How does this compare to a sparkstation > or a Sun 3? MIT people can get an 8 MEG NeXTstation with 105MB and software > for $3450. This isn't terribly different than what we can get for a > sparkstation, which is is said to have 15.8 Dhrystone MIPS, and 1.7 MFLOPS > double-precision LINPACK. I get the idea Sun is leveling with us, not > just telling us maximum theoretical speed we can never acheive, but the > .... I ran double-precision LINPACK benchmark on both NeXTstation and Sparc 1+. I got 1.25 Mflops on Sparc with the cc supplied, 1.31 Mflops with gcc on Sparc and 1.53 Mflops on NeXTstation with the gcc supplied. Although NeXTstation is quite fast in doing floating point multiply and divide, it is very slow doing the operations not implemented in 68040 hardware like log, exp, log, sin, cos etc. These operations generally ran 50-100% slower than 25MHz 68882 and 2-3 times slower than Sparc 1+. I wonder if this is due to some bugs in NeXT operation system because Motorola claimed 25MHz 68040 is supposed to do all those operations faster than 33MHz 68882. ------------ Bing Chen NeXT mail: bchen@pooh.caltech.edu