Computer Speed Claims 1980 to 1996
By Roy Longbottom
This page was set up as 770 pixels wide and accommodates preformatted text <PRE> results tables. Some browsers
produce monospaced font of an unexpected size but this might be adjustable via browser Preferences.
This document contains performance claims and estimates for more than 2000 mainframes,
minicomputers, supercomputers and workstations, from around 120 suppliers and produced between 1980 and 1996.
Speed is given as MIPS (Millions of Instructions Per Second). Maximum MFLOPS
(Millions of Floating Point Instructions Per Second) are also provided,
usually for systems with supercomputer vector processing capabilities.
Where available, production dates and cost is also shown. Details of IBM’s
larger mainframes and PC CPUs up to 2004 have also been included.
Go To Supplier Index
More Historic Data
In 1980 I created a database to contain information about mainframes,
minicomputers, supercomputers and workstations. The information came
from manufacturers’ publications and press reports. Performance numbers
are claims, based on benchmarks or manual calculations. They are not
likely to be particularly accurate but are probably suitable for historic
comparisons. Further data was added until 1996 and lately for Intel and AMD CPUs
(based on Dhrystone benchmark results) and IBM’s larger mainframes up to
2004. The basic data for the latter was found at:
Details provided are Manufacturer, Processor/System Name, Number of
CPUs, CPU/Technology Type, CPU MHz, Operating System, MIPS, Maximum
MFLOPS, Year, Cost and System Type using the following codes:
AR Array Processor
FT Fault Tolerant Server
SF Small Mainframe
VC Vector processing Chip
VU Vector processing Unit
VL Very Long Instruction Word
DB Database Server
Maximum MFLOPS shown are mainly for systems with supercomputer/vector
processing capabilities. These processors can generally produce one on
more results per clock cycle where, for example, at two results per
clock, maximum MFLOPS is two times CPU MHz. Note that modern PCs, using
the new data streaming instructions, can also produce one or two
floating point results per clock cycle.
MIPS figures shown may be based on calculations or benchmark results
and different methods are likely to be used by the various computer
suppliers. The speeds generally only represent raw processing speed and
usually bear no relationship to system throughput capabilities. An
exception could be some mainframes where careful attention is paid to
balanced configurations such that MIPS for a particular range of systems
remains proportional to throughput.
MIPS for many UNIX based systems and PCs are derived from the Dhrystone
benchmark. This uses little data, making it independent of cache sizes (let alone memory size/speed).
This benchmark (and others) may be over-optimised by the compiler to
produce misleading results and some have been excluded where this was
obvious. At least those for Intel and AMD CPUs shown are based on the
same compiled code (one of my benchmarks).
Even if the results represented average processing speed, architectural
improvements lead to wider variations on different applications.
In comparing systems it should be accepted that any MIPS figure might
be 200% of a more appropriate rating, making them unsuitable for use in
competitive purchasing. However, for historical purposes, on comparing
systems of different ages, it does not matter that much.