PONDERING PERFORMANCE
---------------------

Chris-

This is in response to your article "IBM GETTING 190 SPECMARK IN THE
LABS, BUT MICROSOFT EXCEL RUNS SLOW" in the last PowerPC Newsletter. I
am in charge of competitive benchmarking for IBM's Power Personal
Systems Division and thus I either measured or had measured all of the
performance numbers to date on the Power Series machines (desktops and
notebooks) as well as the RS/6000 43P. I'd like to clarify a few of the
points you addressed and emphasize a few others.

In short while our SPEC story is very impressive in terms of price/
performance (I'll address the 200 SPECint vs. 176.4 SPECint question
below), the PowerPC performance advantage truly comes out when you look
at real applications running on our machines under PC-space operating
systems like Windows NT.

Let me try to put the SPECint issue to rest: The 43P delivers 176.4
SPECint in a configuration (133 MHz 604, 32M memory, 512K L2 cache) that
will cost about $7000, an excellent price/performance story. In
announcing the availability of the 604 processor at 133 MHz the IBM
Microelectronics s Division (IMD) supplied the 200 SPECint figure as
their estimate of maximum chip performance. IMD is in the process of
building a "hot box" with larger L2 cache and larger and faster memory
to deliver this figure, much in the same way Intel used an expensive
system with 1M L2 cache, 64M memory, and a separate disk controller with
its own 4M cache to achieve the 155.5 SPECint they announced for the 133
MHz Pentium.

Where it gets sticky is that the high volume Intel machines from the
likes of Dell, Gateway or Micron don't play in the workstation market so
nobody has disclosed a SPEC92 figure on one of these machines. Our
machines *do* play in that market so we measure and disclose SPEC92
numbers that come within 12% of the maxmimum chip performance. I'd be
very surprised if the typical high volume Intel machine could come this
close to Intel's hot box figure.

SPEC is an important benchmark for the workstation space but what most
end users care about is application performance. Thus in our lab here
in Austin we run our Power Series machines side-by-side with
best-of-breed Pentium machines running at the same clock speed and bench
them using real world applications under Windows NT (we'll start
benching OS/2 Warp Connect PowerPC Edition this summer). For new,
32-bit programs that were designed from the ground up to take advantage
of the new processors (that effectively utilize the floating point unit,
for example) we are seeing 2X or greater performance on the Power Series
machines vs Pentium machines at the same clock speed. We have seen this
on both CAD applications and business applications.

For older code like the current 32-bit versions of Microsoft Word and
Excel, which are based heavily on the 16-bit Intel versions and don't
make effective use of the FPU (since they had to run on machines that
didn't have one) our performance advantage is less pronounced. The
slowness of 16-bit apps converted to 32-bit is not particular to
PowerPC.
As Jim Seymour pointed out in his column in the July PC Magazine, it
will take awhile before the software vendors fully utilize the 32-bit
power.
I invite all your readers to check out our performance brief in IBM's
Web site (http:/www.austin.ibm.com/pps/p4be1.htm#deskdesc) and to join
our performance discussion in the PowerPC forum on CompuServe (GO:
POWERPC).
Thanks for your time and your continuing interest in the PowerPC in
all its forms.

Dave Jaffe, IBM Power Personal Systems Performance
jaffe@austin.ibm.com