Conference 7.286::digital

    John, thanks for posting this information. I have seen your name
    in other conferences where performance debates are ongoing. I think
    its important for all of us to understand and track the fast paced
    changes occurring in this area.
    
    Steve Datzman
    

Interesting stuff.  The one question that comes to mind is when systems are
compared, are they of a comparable price?  There's little value in a 100K
system outdoing a 50K one, but if for the same (or better still less) $$$ we're
kicking butt, that's worth mentioning.  Then again, to people who are familiar
with all the different systems in the chart (which I'm not), perhaps the price
positioning is already well understood.

-mark

    I have a question about how SPECs are arrived at.  Are they not
    operating system dependent?  What I am getting at is are the HP marks
    gotten with a 32 OS and ours with a 64 bit OS.  If this is the case,
    what would their marks be when they have a true 64 bit OS?  And what
    will the effect be on ours for NT when 64 bit NT ships?
    
    Excuse the novice questions please.

	The spec specint95 and specfp95 benchmarks cannot exploit
 extended addressing or large physical memories.  Do not expect the
 performance results to change significantly on a platform by virtue of 
 whether they are run in 32bit or 64bit mode.

	So, I do not expect hp's results to get better by using 64bits
 and I do not expect our nt results to improve when ms adds 64 bit features to
 wnt.

	Please note that my commetns apply to the current spec95 suite - there
 is nothing to prevent the spec group to add benchmarks in the future that
 do differentiate based upon wordsize.

	regards,

	Mike

---------------------------------------------------------------------------
This memo is from Rich Grove (compilers) and John Henning (CSD Performance)
---------------------------------------------------------------------------

The CSD Performance Group has submitted new SPEC95 results for most
AlphaStations, with a median improvement of 12% on SPECint95 and 8% on
SPECfp95.  This improved performance is due to software changes in:

       - DEC C 
       - Digital Fortran
       - GEM
       - om
       - KAP for Fortran
       - tuning and performance enhancement efforts in CSD/PG

Thanks are due to all the above-named groups, plus the SEG/AD group, a
consistent source of excellent performance advice.

The chart that follows is sorted by SPECfp95. "Old" indicates performance as
announced between August 1995 and March 1996.  "New" indicates performance
results submitted to SPEC between May 13, 1996 and July 8, 1996.  

           AlphaStation     ------SPECint95-----     -------SPECfp95------  
           Model            Old     New   Change     Old      New   Change  
           ---------        -----  -----  ------     -----   -----  ------  
           200 4/100         1.48   1.88   27%        2.43    2.79   15%    
           200 4/166         2.31   2.95   28%        3.22    3.64   13%    
           200 4/233         3.39   4.28   26%        3.95    4.32    9%    
           255/233           3.80   4.27   12%        5.09    5.09    0%    
           255/300           4.51   5.23   16%        5.71    5.81    2%    
           250 4/266         4.18   5.18   24%        5.78    6.27    8%    
           500/266           7.29   7.93    9%       10.5    11.1     6%    
           600 5/266         6.43   7.91    9%       11.2    11.8     5%    
           500/333           8.87   9.82   11%       11.6    12.5     8%    
           600 5/333         9.23   9.78    6%       13.2    13.4     2%    
           500/400         >11     12.3    12%      >13      14.1     8%    

A graph of the above data may be found in

       PERFOM::CSG_REPORTS:SPEC95_CHANGES_960710.PS

and a Performance Flash additional details will be placed in the
CSG_REPORTS directory on PERFOM:: by July 17.

The remeasurement effort was approved by Workstations Product Management.
The AlphaStation 200 and 250 were included due to existing government
contracts (see June 11 Sales Update).

Finger problem somewhere???

           500/266           7.29   7.93    9% ( 8.77)      10.5    11.1     6% 
  
>>>        600 5/266         6.43   7.91    9% (23.02)      11.2    11.8     5% 
  
           500/333           8.87   9.82   11% (10.71)      11.6    12.5     8% 
  

                                               ^^^^^^^

Gerald

    Finger?  No, not quite... it's just that this one pager was a
    condensation of an 11 page memo with lots of details that I dumped
    during the condensing.
    
    Here's the relevant original part, together with the asterisk that I
    dropped in the summary:
    
-------------------------------------------------------------------------------
                             AlphaStation 600

          --------SPEC Web Page-----------  Digital  CSG_     ------New----
Model     Date Cache Metric          Speed  Web      Reports  Speed %change
--------- ---- ----- --------------  -----  -------  -------  ----- -------
600 5/266 7/95 2mb   SPECint_base95    6.30
                     SPECint95         6.30
                     SPECfp_base95    10.0    10.0   10.0
                     SPECfp95         10.5           10.5

          1/96 2mb   SPECint_base95    6.32    6.32  6.32
                     SPECint95         7.29          7.29

          7/95 4mb   SPECint_base5     6.43                    6.57     2%
                     SPECint95         6.43                    7.91     9% *
                     SPECfp_bas95     10.6                    11.0      4%
                     SPECfp95         11.2                    11.8      5%

                                             * 23% vs. original 4mb submission,
                                                9% vs. 2mb January submission
-------------------------------------------------------------------------------

That is, the AlphaStation 600 5/266 comes in two models: 2mb cache and 4mb
cache.  When we submitted results to SPEC in July/August 1995, we submitted
both 2mb and 4mb, with 6.30 and 6.43 SPECint95 respectively.  In January
1996, we updated the 2mb model (to 7.29), and in July 96 we updated the 4mb
model (to 7.91).

The improvement in July 96 is 9% if you compare it to the most recent
AlphaStation 600 5/266 submission, but is 23% if you compare it to the most
recent 4mb submission.

Thanks for asking, and I'm sorry that I didn't make this point more clear
in the summary.

	/john

    Wow.  Good eyes.
    
    
    Here's another way to look at it, over the past year, our results have
    looked like this:
    
    	AlphaStation	Jul-95			Jan-96			Jul-96
    	600 5/266(2MB)
    	SPECint95	6.30			7.29	       		-
    
    	AlphaStation	Jul-95			Jan-96			Jul-96
    	600 5/266(4MB)
    	SPECint95	6.43			[7.29e]			7.91
    
    
    	AlphaStation		Nov-95	Dec-95	Jan-96			Jun-96
    	600 5/333
    	SPECint95		[8.11]	[8.76]	9.23			9.78
    
    Numbers in []'s were never published.  You can see a big jump in our
    performance early this year as we began incorporating new compiler
    technology (including feedback), new run time libraries, and of
    course new OS.  And of course faster Alphas.  The improvements have
    continued up through shipping the new software and hardware.  (Unlike
    some of our competitor's who have had to say "Oops, not as fast as we
    said it was going to be, sorry about that.")  And of course faster Alphas.
    
    
    So, twelve month improvement for an individual workstation like
    the AlphaStation 600 5/266(4MB) is 23%, and the six month improvement
    is more like 9%.  (I know someone will do the math and predict when
    the AlphaStation 600 5/333 will break 10.0 SPECint95.  Please don't.)
    
    
    
    Or still another way of looking at it.  In the past year, the measured
    SPECint95 of our fastest shipping AlphaStation has just about doubled.
    
    (From AlphaStation 600 5/266(4MB) = 6.43 to AlphaStation 500/400 = 12.3.)
    
    								-mr. bill

    How do these new figures compare to our competitors?

    There's two figures you should be concerned with: SPECetc and # of
    systems sold.  We have the highest SPECetc and the Lowest # of systems
    sold.  I'd rather have a # of systems sold than SPECetc.  Our SPECetc
    are better than HP and the New Sparc Ultra but those systems are easier
    to sell.  There's a VAR here that receives about 20 Sparc with every
    UPS shipment morning delivery and re-ships the same units with the
    afternoon UPS pickup.  They also carry Alpha's but demand isn't too
    great for these products.
    Congradulation to Engineering and the Performance Group, now who's
    responsible for taking this information and create the demand for the
    product?

    re: .10
    
    Two things about Spec etc and #'s sold:
    
    1) Re: "Our SPECetc are better ..."
    
    We are not better than HP in all Spec rates.
    
    2) Re: #'s sold.
    
    The SPEC rates are helpful for getting units sold and, thus, we should
    be concerned.