[gentoo-user] [OT]Real compute cycles in a laptop?

[gentoo-user] [OT]Real compute cycles in a laptop?

[Mark Knecht]

Hi,
   I'm in the market for a new laptop. I sold mine about a year ago
and it's time to get a new one. I will run both Linux and Win XP on
the machine. The Linux workload will not be anything substantial but
when in Win XP I have a program that I'll be running that is very
compute bound on my only XP desktop machine.

   Does anyone know of good comparative data on BogoMIPs in Linux vs.
real compute speed for different processors? I'm looking for some way
to compare different processors in the laptops I'm looking at vs. my
current XP desktop which is my slowest machine.

   The application under Windows is doing neural network stuff. I have
no idea how much of it is floating point based but my assumption is
that is a pretty big part of the whole picture. Is the AMD FPU still
superior to the Intel FPU or are they at parity these days?

   The current machine has 768MB. The application never uses more than
256MB and there is no significant disk I/O but the processor sits at
100% in XP for hours doing it's work optimizing the neural network.

   Thanks in advance for any pointers.

Cheers,
Mark
-- 
gentoo-user@gentoo.org mailing list

Re: [gentoo-user] [OT]Real compute cycles in a laptop?

[Alan McKinnon]

On Tuesday 07 November 2006 05:18, Mark Knecht wrote:
> Hi,
>    I'm in the market for a new laptop. I sold mine about a year ago
> and it's time to get a new one. I will run both Linux and Win XP on
> the machine. The Linux workload will not be anything substantial but
> when in Win XP I have a program that I'll be running that is very
> compute bound on my only XP desktop machine.
>
>    Does anyone know of good comparative data on BogoMIPs in Linux vs.
> real compute speed for different processors? I'm looking for some way
> to compare different processors in the laptops I'm looking at vs. my
> current XP desktop which is my slowest machine.
>
>    The application under Windows is doing neural network stuff. I
> have no idea how much of it is floating point based but my assumption
> is that is a pretty big part of the whole picture. Is the AMD FPU
> still superior to the Intel FPU or are they at parity these days?
>
>    The current machine has 768MB. The application never uses more
> than 256MB and there is no significant disk I/O but the processor
> sits at 100% in XP for hours doing it's work optimizing the neural
> network.

You seem to know enough about these matters already to make a sane 
judgement, so you probably already know that the real answer to your 
question is "it depends".

Here's what I would do: pop along to your local store, preferably not 
one of the big chains, find the sanest sales guy with a clue and 
explain your problem. Don't listen to his recommendations, just ask if 
you can test his demo machines with the actual app in question. If it's 
an owner run store he probably say yes. Then test the thing for real 
and measure progress after 30 minutes or so. Buy the best performer.

This will take a while, but at least you'll know for real which one 
suits your needs best

alan
-- 
gentoo-user@gentoo.org mailing list

Re: [gentoo-user] [OT]Real compute cycles in a laptop?

[Mark Knecht]

On 11/7/06, Alan McKinnon <alan@linuxholdings.co.za> wrote:
>
> You seem to know enough about these matters already to make a sane
> judgement, so you probably already know that the real answer to your
> question is "it depends".
>
> Here's what I would do: pop along to your local store, preferably not
> one of the big chains, find the sanest sales guy with a clue and
> explain your problem. Don't listen to his recommendations, just ask if
> you can test his demo machines with the actual app in question. If it's
> an owner run store he probably say yes. Then test the thing for real
> and measure progress after 30 minutes or so. Buy the best performer.
>
> This will take a while, but at least you'll know for real which one
> suits your needs best
>
> alan

Thanks Alan.

The problem with running the neural network app is that it's a huge
install under Windows. It requires Internet access as it has a
hardware key that has to be validated against the specific machine.
Probably takes 1 hour just to set up. Then, once it's set up it takes
maybe 15 minutes to run a single solution on my older Athlon XP 1600+.
With that as background I'm sure you can understand that I'm not
anxious to do it more than once or twice.

What I was hoping to do was find some basic way of comparing the
BogoMIPS on my old Athlon XP machine with BogoMIPS on some new
machines at the dealer. They haven't had any problems in the past with
me bringing in a LiveCD and booting Linux. If I could do this then I
might estimate that the new machine will run the same speed or will
run 3X the speed when doing these neural network jobs?

Here's some info on the machines in my house today:

1) A 3GHz P4HT machine we use as a MythTV backend server and desktop machine:

processor       : 0
vendor_id       : GenuineIntel
cpu family      : 15
model           : 3
model name      : Intel(R) Pentium(R) 4 CPU 3.00GHz
stepping        : 3
cpu MHz         : 2995.346
cache size      : 1024 KB
<SNIP>
bogomips        : 5996.11

processor       : 1
vendor_id       : GenuineIntel
cpu family      : 15
model           : 3
model name      : Intel(R) Pentium(R) 4 CPU 3.00GHz
stepping        : 3
cpu MHz         : 2995.346
cache size      : 1024 KB
<SNIP>
bogomips        : 5990.25

2) My son's AMD Compaq low cost machine:

processor       : 0
vendor_id       : AuthenticAMD
cpu family      : 15
model           : 47
model name      : AMD Sempron(tm) Processor 3200+
stepping        : 2
cpu MHz         : 1803.767
cache size      : 256 KB
<SNIP>
bogomips        : 3611.84

3) 1 of 2 Pundit-R's used as Myth frontend machines:

processor       : 0
vendor_id       : GenuineIntel
cpu family      : 15
model           : 3
model name      : Intel(R) Celeron(R) CPU 2.26GHz
stepping        : 4
cpu MHz         : 2261.847
cache size      : 256 KB
<SNIP>
bogomips        : 4526.57


4) My AMD64 Gentoo machine is use daily:

processor       : 0
vendor_id       : AuthenticAMD
cpu family      : 15
model           : 47
model name      : AMD Athlon(tm) 64 Processor 3000+
stepping        : 0
cpu MHz         : 1809.286
cache size      : 512 KB
<SNIP>
bogomips        : 3619.63

5) The current Athlon XP Windows machine is busy running Trading
Solutions but when in Linux I *think* it has a BogoMIPS spec around
2800. No way to verify that right now.

It's pretty boring but it seems that you can sort of double the CPU
MHz spec and come pretty close to the BogoMIPS numbers. However that
doesn't take cache size into account so maybe BogoMIPS isn't even the
right thing to be looking at.

At the root of it all my questions are:

1) How representative are BogoMIPS in determining how fast a machine
will eventually be on compute bound apps?

2) Are there any good listings of BogoMIPS on different processor
types and speeds? (This is what I really want....)

3) Do BogoMIPS include FPU measurements in case that is important for my app?

4) How well do BogoMIPS translate to the same machine when it runs Win XP?

Anyway, thanks very much for your answers. I appreciate the help even
if it isn't primarily about running Gentoo. I suspect there are others
out there that have to dual boot. Maybe this will help someone in the
future with similar questions.

Cheers,
Mark
-- 
gentoo-user@gentoo.org mailing list

Re: [gentoo-user] [OT]Real compute cycles in a laptop?

[Alan McKinnon]

On Wednesday 08 November 2006 19:00, Mark Knecht wrote:

> The problem with running the neural network app is that it's a huge
> install under Windows. It requires Internet access as it has a
> hardware key that has to be validated against the specific machine.
> Probably takes 1 hour just to set up. Then, once it's set up it takes
> maybe 15 minutes to run a single solution on my older Athlon XP
> 1600+. With that as background I'm sure you can understand that I'm
> not anxious to do it more than once or twice.

OK, scrap that idea

> What I was hoping to do was find some basic way of comparing the
> BogoMIPS on my old Athlon XP machine with BogoMIPS on some new
> machines at the dealer. They haven't had any problems in the past
> with me bringing in a LiveCD and booting Linux. If I could do this
> then I might estimate that the new machine will run the same speed or
> will run 3X the speed when doing these neural network jobs?

For this purpose bogomips is meaningless - it's simply a measure of how 
fast the cpu can execute a very specific and very tight loop and is 
used for some timing setting or other during kernel initialization. 
Bears almost no resemblance to real life operation, other than it's 
safe to say that bogomips goes up as cpu freq goes up

[snip]

> 5) The current Athlon XP Windows machine is busy running Trading
> Solutions but when in Linux I *think* it has a BogoMIPS spec around
> 2800. No way to verify that right now.
>
> It's pretty boring but it seems that you can sort of double the CPU
> MHz spec and come pretty close to the BogoMIPS numbers. However that
> doesn't take cache size into account so maybe BogoMIPS isn't even the
> right thing to be looking at.

bogomips is usually about double the cpu speed, but you can't count on 
that. I would imagine that cache size and fpu performance were 
significant factors.

Let's assume this app of your is floating point intensive (fairly safe 
assumption), and doesn't use a heck of a lot of RAM or disk (already 
shown to be true). So now you need to rate the fpu of the various 
processors and machines around. So I would read reviews of various 
machines in computer performance mags where they publish sane 
benchmarks, to get an idea of what would be best

alan
-- 
gentoo-user@gentoo.org mailing list

Re: [gentoo-user] [OT]Real compute cycles in a laptop?

[Mark Knecht]

On 11/8/06, Alan McKinnon <alan@linuxholdings.co.za> wrote:
<SNIP>
> > It's pretty boring but it seems that you can sort of double the CPU
> > MHz spec and come pretty close to the BogoMIPS numbers. However that
> > doesn't take cache size into account so maybe BogoMIPS isn't even the
> > right thing to be looking at.
>
> bogomips is usually about double the cpu speed, but you can't count on
> that. I would imagine that cache size and fpu performance were
> significant factors.
>
> Let's assume this app of your is floating point intensive (fairly safe
> assumption), and doesn't use a heck of a lot of RAM or disk (already
> shown to be true). So now you need to rate the fpu of the various
> processors and machines around. So I would read reviews of various
> machines in computer performance mags where they publish sane
> benchmarks, to get an idea of what would be best
>
> alan

Yeah - makes sense. Thanks.

Back to Google FPU stuff.

Thanks for your inputs.

Cheers,
Mark
-- 
gentoo-user@gentoo.org mailing list