| Re: Platform recommendation/tradeoffs? [message #3953] |
Fri, 14 April 1995 00:00  |
foresto
Messages: 1
Registered: April 1995
|
Junior Member |
|
|
In article <3mhc42$6lt@mozo.cc.purdue.edu>, smithjp@sage.cc.purdue.edu (Joseph Smith) writes:
|> I am about to order some flavor of IDL (Mac or Unix) and
|> I'm wondering if anyone can help me gauge just how much performance
|> I loose by getting the Mac product. (This saves me enough money
|> to buy a Q630 cpu)
|>
|> I have a SG Iris Indigo with lots of memory and some disk space
|> that I could be using for IDL, or I could put it on the aforementioned
|> Q630. If intense image processing were among my requirements it would
|> be a pretty easy decision, but the data that I'll be working on is
|> basically about 16 or 32 128x128 matrices - not too big of a problem
|> from a graphic standpoint (I suppose I would rarely need to look at
|> more than 1 128x128 image at a time, for instance)
|>
|> This problem has a lot of number crunching - 128x128 non linear
|> least squares fits of at least 4 parameters per fit.
|>
|> Any comments? Any experience using IDL on a 68040?
|>
I ran a small testbench between a SPARC20, a PowerMac and a Quadra. The test
included computing a Fast Fourier Transform on a 1024x1024 float array, and
1000 times computing a 1024 FFT. Here are the results:
Machine FFT 1024x1024 1000 FFT 1024
Sparc 20 8.5s 3.4s
PowerMac 7100/66 14.6s 7.4s
Centris (68040@25MHz + FPU) 39.8s 33.4s
A few comments:
1- Take it for what it's worth. It's only a **vague** indicator of raw processing
power. I does not tell anything about disk I/O, where the Unix machine outperforms
the Mac several times. The Mac OS file manager is notoriously slow (it is still
emulated), but this should improve in the future version of the system to be
delivered in 1 year (does anyone have precisions or more quantitative information
on this?)
2- The Mac and PowerMac are running demo versions of IDL (which refuses to run
the built-in testbench (run in 39s by the Sparc20): I don't know why this feature
has been inhibited on demo versions btw).
The PowerMac is running "native" code, however it's not that much faster than
the Centris. I suspect most of the "native" code is actually still emulated. On
FPU intensive operations a PowerMac running real native code is usually about
8 times faster than a 25MHz Quadra. All machines here were running IDL 3.6 with
sufficient memory (>25Mb) for the test.
3- Notice how the ratio of the times needed to perform the two test benches
changes with the processor. Does anybody have a clue?
Also, I should say that the "look and feel" of IDL for the Mac is surprising
to say the least: buttons do not have the standard Mac appearance, windows are
most often not resizable, etc... Clearly IDL is not a Mac application, it's an
application from the mainframe world that has been ported to the Mac, with
minimal changes. On a 32768 color monitor, I was not able to get allocate 256
colors for IDL while keeping 3D-looking system windows in the background.
One can only hope that most of these non-Mac quirks will
disappear with version 4.0. Does anyone know how IDL 4.0 for the Mac will
look like? Will it be any faster than 3.6 on PowerPC machines?
Another Mac-related question: does anybody know if it's possible to produce
"real color" pictures (from green-red-blue frames for example) on a Mac with
a 12 bit (or more) color monitor?
My conclusion: if you are doing only data crunching, use few widget applications
(which look ugly on the Mac, what an irony!), and do not want to spend too much
money, then probably your best deal would be to buy a PowerPC Mac 6100/66. I
would strongly discourage buying a Quadra, even if that means saving a few
hundred bucks. IDL and the MacOS are going to run faster on the Power Macs as
more and more of the code runs native, while it might be very well possible
that RSI will discontinue IDL for 680x0 CPUs in a few years. If your application
has intensive file I/O, then you'd better stay with a Unix machine. To give an
order of magnitude, loading a 256x256x10 data cube is unpractically long on the
Mac for the moment (maybe 1min or more, against about 10-15s on the Sparc20).
Vincent Foresto
Max-Planck Institut fuer Astronomie, Heidelberg
|
|
|
|
| Re: Platform recommendation/tradeoffs? [message #3955 is a reply to message #3953] |
Fri, 14 April 1995 00:00  |
kennealy
Messages: 11
Registered: August 1994
|
Junior Member |
|
|
daft@debussy.crd.ge.com (Chris Daft) writes:
> I second the request for some comparison of IDL's performance on
> various CPUs. I have been using IDL on Suns for a long time, but may
> be required to switch to a Pentium machine for some period, and I'd
> like to know what the performance differences would be.
> I wonder if some kind of standard IDL test suite could be developed,
> kinda like a SPECmark?
Every IDL distribution includes a copy of time_test.pro, which is
exactly the standard test suite you refer to. I've run this on quite
a few different machines, but I don't have the data here right now
in my home office. However, I do recall that my 90MHz Pentium system
fares somewhat better than a a SUN SPARCstation-10 on this test. I
was a little surprised by that result, so I used one of my own
Fortran FFT routines to compare my Pentium with the SPARC-10 and
got pretty much the same result. The Pentium is indeed impressive.
--
============================================================ ===
Dr. Jack Kennealy, Nashua, NH kennealy@mv.mv.com
============================================================ ===
|
|
|
|
|
|
|
|
|
|
|
|
| Re: Platform recommendation/tradeoffs? [message #4047 is a reply to message #3953] |
Sat, 15 April 1995 00:00 |
kennealy
Messages: 11
Registered: August 1994
|
Junior Member |
|
|
foresto@sun16.NoSubdomain.NoDomain (Vincent Coude' du Foresto) writes:
> I ran a small testbench between a SPARC20, a PowerMac and a Quadra. The test
> included computing a Fast Fourier Transform on a 1024x1024 float array, and
> 1000 times computing a 1024 FFT. Here are the results:
> Machine FFT 1024x1024 1000 FFT 1024
> Sparc 20 8.5s 3.4s
> PowerMac 7100/66 14.6s 7.4s
> Centris (68040@25MHz + FPU) 39.8s 33.4s
Add this data to the table
90MHz Pentium (32 MBytes RAM) 10.3s 3.2s
> the built-in testbench (run in 39s by the Sparc20)
On the Pentium, this runs in 17.3s.
--
============================================================ ===
Dr. Jack Kennealy, Nashua, NH kennealy@mv.mv.com
============================================================ ===
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
