discrete transistors - test-drb@ccmp.vtda.org

21 Oct 1998

...
 >Irrelevant to the discussion for two reasons:
>1)  No one has made mainframes out of discrete transistors since the late 60s.
>2)  Almost all modern mainframes are based on monolithic microprocessors. 
...
  My point was that only recently has it become so that
there is nothing 
 more than an Alpha or an PII. There are many parallel implementations, 
 but nothing like the difference between an Altair and an S/360. Yes, I 
 know there are many computers way more powerful than CompUSA desktops, 
 but the point is that they're just parallel implementations and there 
 isn't so much of a class distinction. 
Do you want to tell us that the mainframe world
didn't evolve ? Boy, they also have new and quite
speedy new processors - for example, the new SIEMENS
S150 delivers around 93 RPF on a single CPU and 
a sustained rate of 760 RPF on a 12 CPU configuration.
RPF (Relative Performance Factor) is a SIEMENS unit
to classify the relative speed of their processors.
The baseunit is a processor from the beginnign of the
80's. Even in the main frame (/390 compatible) world
it's hard to compare processors just by speed, since
the manufacturers use different architectures below.
So every one has his own measurement. IBM uses their
MIPS (misleading informations on processor speed) as
a guideline - but even these are just a definde rate:
they just fixed the 9672-R15 as base with 63 MIPS.

A rough conversation between IBM MIPS and SIEMENS RPF
is 1.3 to 1.5 - so die S150 deliveres between 140 and
1000 IBM MIPS sustained load at average mix.

With the small SR2000 systems which are based on
MIPS (now thats the processor manufacturer :) R10000
CPUs there is a guideline for transformation: one
R10000 deliveres around 12 RPF or 15 MIPS at user level.
(if you know take the MIPS R10000 MIPS rating you'll
get also a fairly nice RISC to (IBM)CISC ratio, since
the firmware interpreter used in the SR2000 is very
performant)

Now if you take the SpecInt value of the R10000 and
of the Alpha or Pentium chips (sorry, I just not
have them ready), one could base a comparisation on
these values. If a Pentium deliveres about 10 times
the performance of a R10000, they might reach te region
of an actual mainframe (not any 60's transistor claim).

But in fact these numbers are again way beside the
actual usability - Mainframe performance is measured
by thruput and things like transactions per second
within a defind response time rather than raw CPU.

Just take a look at http://www9.s390.ibm.com/lspr/
to get an idea how specific measurment in the mainframe
world is don - not just instructions or megabytes
per second.

Gruss
H.

--
Ich denke, also bin ich, also gut
HRK