# Re: Singularity and AIs (was: We're stuck with each other)

Date: Sun Jan 27 2002 - 19:56:50 MST

On Sun, 27 Jan 2002, Party of Citizens wrote:

> Can anyone explain the difference, if any, between an operation per
> second, a calculation per second and an instruction per second?

It depends on the nature of the hardware. As Dan indicated, "FLOP"
which stands for floating point operation per second has a pretty
standard definition. The design of most floating point units usually
allowed the ADD, SUB and MULT operations to be carried out in the
same period of time. This used to be in 1 clock cycle, though I think in
most CPUs today they take several cycles. DIV typically took a bit
longer. I don't know if anyone ever designed a floating point unit
where they put enough hardware into it that the DIV executed in one
clock cycle as well (Google might know...).

Now on the older hardware particularly the difference between an
instruction per second and a (Floating) "operation per second"
was significant. On PDP-11's the clock cycle time was 300 ns
and that was how long it took for most integer instructions
(int-ADD, SUB, MOV, etc.). So if you were just executing integer
code, you could get 1/3 MIPS out of the CPU. Floating point
operations took several cycles, so you usually got more like
less than 1/6 MFLOPS. So "IPS" were not generally equal to "FLOPS".
On modern CPUs I think there is enough floating point hardware
and the CPUs are so pipelined that I think the difference is
less pronounced.

Now "calculations per second" gets into the nature of the machine.
For general purpose computers, its just IPS or FLOPS, but in
special purpose computers it is more complex.

Chesire I think asked what a GRAPE was -- that a special purpose
"GRAvity PipelinE" computer that was designed by a group in Japan
to do just the gravity calculation. It did the calculation in hardware
and was massively parallel, There have now been 5 or 6 generations of them,
with each generation being faster. It is used to simulate things like dust
clouds collapsing into stars, billions of stars collapsing into galaxies,
two galaxies colliding with each other, etc. Since the fundamental execution
unit is the entire gravity calculation (I'm unsure of how many MULTs/ADDs/EXP, etc.
this is) -- comparison between GRAPEs would be measured in GrvtyCalcsPS.
However, I think they generally figured out how many FLOPS this was
equivalent to and so they "rated" GRAPES like other computers (GFLOPS & TFLOPS).
For most of the last decade they've been the most powerful computers on
the planet in terms of raw throughput (though for a very special
application).

Robert

This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 13:37:36 MST