Conference ricks::dechips

    
    >how do i compare them??
    
    They don't compare very well.  SPEC95 results were generated for
    a Sparcstation [45]/110, with 128Mb memory, 1.5Gb disk, running
    Solaris 2.5.1 beta (according to http://www.specbench.org/)
    With C and FORTRAN programs.
    
    The microsparc has a floating point unit on board.  The SA110 (the
    CPU in our NC reference design) has none.
    
    The original 'IT' (Internet Terminal) board, on which the caffeinemarks
    were measured, has 8Mb memory (not adjustable), NO disk, and was
    running JavaOS (also a beta).  A FORTRAN compiler for this machine
    doesn't exist.  And even if it did, a number of the SPEC95 programs
    simply wouldn't run, because the code size + memory size of the runs
    exceeds 8Mb.
    
    The DNA (Digital Network Appliance) can be expanded all the way
    up to 64Mb of memory.  It does not have a disk, either.
    There is no FORTRAN compiler for the DNA (or the IT)
    
     I haven't seen any performance numbers on Javastations (let alone
    any Javastations - I don't know if they are actually shipping).
    Engineering people familiar with the DNA have claimed that the
    Javastation demos they have seen have all been very slow.
    
    I don't think you will see any apples-to-apples comparisons on the
    various NCs until SPEC has a Java suite which is designed to run
    on memory limited machines.
    
    Until then, I think caffeinemarks (and some other Java benchmarks)
    are going to be the best ways to compare NCs.  I'm not sure how much
    memory a JIT consumes, so it is unclear whether or not doing a
    performance comparison between Java interpreter numbers and JIT numbers
    is meaningful.  (i.e. if a JIT requires 8Mb memory to generate results,
    then it clearly isn't meaningful in a non-bloated NC).
    
    Sorry if this doesn't clarify things, but there simply isn't a good
    comparison.
    
    	Rick
    
    Richard Gorton                  All standard disclaimers apply.
    Alpha Migration Tools           Projects: DECmigrate (mx), FreePort Express,
    Digital Semiconductor                     Linux/Alpha, JavaOS
    Digital Equipment Corporation  
    http://www.digital.com/info/semiconductor/amt
    Reply-to: gorton@tallis.enet.dec.com

    i read on www.eu.sun.com/sparc/whitepapers/wpr-0015-01 
    that a compiler (25.000 lines of code) is about 450k java Bytecode
    i read that there some where there is a java version of Spec suite.
    
    i havent' found any caffeinemarks for Javastations...
    if anybody does, pls let me know..
    did you get a javastation?
    
    
    Brunoi

When SPEC95 first came out, a number of machines were measured with both
SPEC95 and SPEC92.  Their SPEC92 ratings were about 35x to 45x their SPEC95
ratings, if I recall correctly.  SPEC92 was defined using the VAX-11/780
as the basis.  So as a wild extrapolation, you could estimate that the
MicroSparc 2 has integer performance of about 64x a VAX-11/780, and floating-
point performance of about 80x a VAX-11/780.  Now, recall that "Dhrystone
MIPS" were based on calling a VAX-11/780 a 1-MIPS machine, and you have the
basis for a rough comparison.  Remember that Dhrystone performs no floating
point, and that the SA-110 has no FPU...

    
    	I found one reference for caffeinemarks on the Javastation
    	it was reported as 127 for Caffeinemarks 2.5 (add 5-10% to 
        estimate Caffeinemarks 2.01).
    
    	At any rate the StrongARM results for @185Mhz and @ 235Mhz 
       SA-110 were 422 and 520 respectively on Caffeinemarks 2.01
    
    
    

From http://www.webfayre.com/pendragon/jpr/jpr049702.html#Digital's StrongARM 
Processor:

Digital's StrongARM Processor 

Digital's StrongARM CPU has posted the highest interpreter scores we have seen. 
While fast RISC workstations achieve scores in the 200-300 range, Digital has 
reported scores in the 400-550 range for its 185MHz and 235MHz SA-110 processors 
(devices that are usually used in handheld or embedded applications). The 
following information is taken from a Digital press release: 

NOTE: These scores are CaffeineMark 2.01 scores (they are approximately, but not 
exactly CaffeineMark 2.5 scores). 

Hardware: Digital IT
CPU StrongARM SA-110 (no L2 cache)
Memory: 8Mbytes total ram of which approx 4Mbytes are Java heap
Os/Browser: JavaOs-preAlpha3/HotJava (interpreted code, no JIT)
Graphics 1024x768 256 color, dumb frame buffer on ISA bus
(no graphics acceleration was used)

Start at http://www.webfayre.com/cm.html and work your way down through "The 
Java Performance Report".

PJDM

This didn't seem to make it through the news-notes gateway,
so Im going to try again from the web access point! Sorry
if it gets here twice.


I just pulled the latest data from www.webfayre.com
the home of CaffeineMarks. Extracting the javastation
and strongarm numbers:

cmark   sieve   loop    logic   string  fp      method  graphics image  dialog

JavaStation (box model)microSparc II, ? cache, 40MB mem, JavaOS/HotJava 1.0
LAR
124     114     106     126     99      106     101     64      330     46

Javastation, 100MHz Microsparc-II, 16kb cache, 16MB mem JavaOS/HotJava  1.0LAR
125     115     124     128     100     108     98      67      329     42

Digital IT,233MHz SA-110, no cache, 8M mem, JavaOs preAlpha-3/HotJava
preBeta-2
562     408     474     482     397     309     273     175     440     1797

WYSE 4300 NC, Strong ARM DEC, no cache, 16M mem, JAVA OS 1.0F/HOT JAVA
274     349     402     411     333     264     245     124     393     105
271     349     402     409     333     264     245     113     393     95
272     349     402     411     334     264     245     157     391     56
273     349     404     411     334     264     246     123     394     93


All of these are interpreting Java.

On the CPU intensive tests (sieve, loop, logic and string) the
performance of StrongARM over microsparc-II can be seen.

The floating point numbers for StrongARM would be expected to be low,
since there is no FPU. I think the number there is reflecting the speed
of the rest of the system.

The graphics numbers are dominated by the speed of filled rectangle
painting (which fortunately is the one thing that is easy to 
accelerate in the video controller on IT :-)

The dialog test basically updates some text boxes lots of times. On the
older JavaOs it didn't bother to update on the screen a rapidly changing
text box (thread priorities) this gives the extreemly high result for
the IT (that is not seen on the Wyse version with later JavaOs). I pointed
this out to the CaffeineMark people at the time and it was flagged in our
results posting -- the caffeinemark is supposed to measure "system"
performance
so I was told to report it.

We are not reporting results for the DIGITAL Network Appliance Reference
Design
yet because we only just have JDK 1.1 running and have not done any
optimisation.
However, the numbers look about where the IT was before we optimised that,
so for interpreting I would expect similar results. Once the graphics
accelerator is used the graphis, image and dialog tests should be better
than on IT.

(The IT is also known as the EB-SIT board from DigtalSemi)

Mark


[Posted by WWW Notes gateway]

I just pulled the latest data from www.webfayre.com
the home of CaffeineMarks. Extracting the javastation
and strongarm numbers:

cmark   sieve   loop    logic   string  fp      method  graphics image  dialog

JavaStation (box model)microSparc II, ? cache, 40MB mem, JavaOS/HotJava 1.0 LAR
124     114     106     126     99      106     101     64      330     46

Javastation, 100MHz Microsparc-II, 16kb cache, 16MB mem JavaOS/HotJava  1.0LAR
125     115     124     128     100     108     98      67      329     42

Digital IT,233MHz SA-110, no cache, 8M mem, JavaOs preAlpha-3/HotJava preBeta-2
562     408     474     482     397     309     273     175     440     1797

WYSE 4300 NC, Strong ARM DEC, no cache, 16M mem, JAVA OS 1.0F/HOT JAVA
274     349     402     411     333     264     245     124     393     105
271     349     402     409     333     264     245     113     393     95
272     349     402     411     334     264     245     157     391     56
273     349     404     411     334     264     246     123     394     93


All of these are interpreting Java.

On the CPU intensive tests (sieve, loop, logic and string) the
performance of StrongARM over microsparc-II can be seen.

The floating point numbers for StrongARM would be expected to be low,
since there is no FPU. I think the number there is reflecting the speed
of the rest of the system.

The graphics numbers are dominated by the speed of filled rectangle
painting (which fortunately is the one thing that is easy to 
accelerate in the video controller on IT :-)

The dialog test basically updates some text boxes lots of times. On the
older JavaOs it didn't bother to update on the screen a rapidly changing
text box (thread priorities) this gives the extreemly high result for
the IT (that is not seen on the Wyse version with later JavaOs). I pointed
this out to the CaffeineMark people at the time and it was flagged in our
results posting -- the caffeinemark is supposed to measure "system" performance
so I was told to report it.

We are not reporting results for the DIGITAL Network Appliance Reference Design
yet because we only just have JDK 1.1 running and have not done any optimisation.
However, the numbers look about where the IT was before we optimised that,
so for interpreting I would expect similar results. Once the graphics
accelerator is used the graphis, image and dialog tests should be better
than on IT.

(The IT is also known as the EB-SIT board from DigtalSemi)

Mark
[posted by Notes-News gateway]

.2:
>    i read on www.eu.sun.com/sparc/whitepapers/wpr-0015-01 
>    that a compiler (25.000 lines of code) is about 450k java Bytecode

This compiler, javac, translates java programs into BYTECODE.  This
is not the size of the just-in-time compilers (jits) for generating
native code from bytecode.

.6:
> The floating point numbers for StrongARM would be expected to be low,
> since there is no FPU. I think the number there is reflecting the speed
> of the rest of the system.

The FP portion of caffeinemark actually spends most of its time doing
array subscripting (is the array null?, get the array bound, check the
subscript, fetch the element (pointer to array), is the array null?,
get the array bound, check the subscript, fetch the float value), and not
floating point computation. Here's the inner loop

   for (int i4 = 0; i4 < 3; i4++)
       for (int j3 = 0; j3 < 3; j3++)
         aad4[i4][i2] = aad4[i4][i2] + aad1[j3][i4] * aad4[i4][i2];

so this has 2 FP operations vs. 4 instances of double subscripting