Re: cpu arch performance
Hello Matthias,
On 24 May 2000, Matthias Hertel wrote:
> Daniel Reuter <reuter@Uni-Hohenheim.DE> writes:
> > On Tue, 23 May 2000, Lee Elliott wrote:
> > > Just something I noticed after setting up an x86 system after running
> > > Debian on m68k.
> > >
> > > The m68k system was an Amiga with an m68060/50MHz which gave a BogoMIP
> > > rating of 99.something. The x86 system is a dual PIII 650MHz system and
> > > it rates as 2600 BogoMIPs. This would seem to imply that on a per MHz
> > > basis, the m68060 is the equivalent of two PIIIs.
> >
> > This is right.
> [...]
> > holds true for 386 and 486). This is why a m68k processor is on a per MHz
> > basis really twice as fast as a x86 processor.
>
> Not really.
>
> It is true that the later 68ks drive at least their integer units at
> twice the external clock rate (don't know about the FPU). And, in
> contrast to Intel's 486DX/2 and later processors, it is the external
> clock rate that Motorola tacks onto the processor name.
It has nothing to do with internal and external clock rates, the issue is
about at which moments in the clock cycle does the processor do a certain
thing. You connect a quartz with for example 33 MHz to a 486DX/2 and the
clock is internally doubled to 66 MHz. But the execution of commands is
still only done when the (internal) clock is at a peak. At a m68k system,
the clock is not internally doubled. It just issues commands at peaks AND
lows of the clock signal. So effectivly, a m68k at 33 MHz would be
(concerning only the processor) just as fast as a 486DX/2 without doubling
the clock-frequency internally. (At least in regard to clock frequency,
there are sure more things, which determine processor speed ;-)
(Got this from a book about microprocessor technique, it's a bit outdated
(1994/95) or so, but the issues involving m68k and i486 are described)
> For example, a Pentium gets
> 2 BogoMips per MHz, and a Pentium Pro get 1 BogoMip per MHz, but in
> most situations a Pentium Pro gets a least as much done per cycle as a
> Pentium (this is a slight understatement). Read the Bogomips HOWTO.
To cite the BogoMIPS-HOWTO:
Hence the BogoMIPS value gives some...and so forth..., but it is way too
unscientific to be called anything but BogoMIPS.
But I think in the special case i386/i486 and m68k it really has
to do with the above mentioned issue.
By the way: a Normal Pentium doesn't get 2 BogoMIPS per MHz. Read the
Bogomips HOWTO ;-)
> If you want a speed comparison, time some software that interests you.
> Compile a kernel. Do some large FFTs. Compare these results to the
> clock rate if you like. Compare to hardware cost if you want more
> meaningful numbers.
I don't think this is a good method of measuring "raw" processor speeds.
As compilation is usually a memory-intensive process, the speed depends
heavily on such things as RAM-size (compiled on the same 486 processor, a
certain kernel-compile took 50 Minutes with 8MB RAM and 20 Minutes with 24
MB RAM).
Sure your method would be OK to determine something like let's say
"motherboard (including RAM) speed" or "overall system speed", but it is
not suited to determine the speed of the processor as such (which is - I
submit - something which doesn't really interest anybody except perhaps
microprocessor engineers;-). Most people want to know about the speed of
their complete system).
Regards,
Daniel
P.S.: one more cite from the BogoMIPS-HOWTO
"Why to pay attention to BogoMIPS:
To see whether your system is faster than mine. Of course this is
completely wrong, unreliable, ill-founded, and utterly useless, but all
benchmarks suffer from this same problem. So why not use it? This inherent
stupidity has never before stopped people from using benchmarks, has it?
:-)
Reply to: