Spice Accelerator Hardware?

[Douglas Mota]

Hi Ken,
Do you know why has LTSpice run faster over a 64 bit OS, even though
it's a 32 bit application? Sorry if it's a stupid question...
Regards!

 

[Ken Smith]

In article <4cmds1te6js35tbg0vcikjf87aoc3p0d3d@4ax.com>,
qrk <SpamTrap@spam.net> wrote:
[...]

RAM, sustained hard drive writing speed, and background tasks (like
virus scanners) can have significant effects on program execution
time.

I don't have any viruses to scan for but on a Windows machine I have seen
virsus software slow disk IO way down.

When comparing processors, you need to normalize the clock rate to
some number, like 1 GHz.

How about I normalize to 3.3GHz? Then I don't have to do any math to
compare the 2 machines.

[ observations on PSpice ..]

1) AMD Athlon solved twice (assuming same CPU clock freq) as fast as
the old P4s, about 1.7x faster than the newer P4s. The PIII ran about
1.3x faster than the old P4.

Since spice tends to take a lot of floating point operations, it is the
speed of that part that matters so these sound reasonable.

2) AMD Athlon and Opteron are very close in execution time on Windoze
2000.

That surprises me a little.

3) Memory speed had very little effect on execution speed on the
PSpice sim we were using.

If the floating point operations take multiple machine cycles this is
reasonable.

4) Hard drive speed had a noticeable effect since a large data file is
written out.

Yes and the type of disk structure also matters too I'll bet.

--
--
kensmith@rahul.net forging knowledge

 

[Helmut Sennewald]

"Douglas Mota" <douglasm@uninet.com.br> schrieb im Newsbeitrag
news:1137066341.683300.280310@g49g2000cwa.googlegroups.com...

Hi Ken,
Do you know why has LTSpice run faster over a 64 bit OS, even though
it's a 32 bit application? Sorry if it's a stupid question...
Regards!

Hello Douglas,

there may be only small differences in speed between all the OS.
Nobody has given any number so far in this discussion. So please
be aware that we talk here about a few percent difference.

We have collected benchmark results of three smaller SPICE-circuits
in the database of the LTspice-Yahoo-group.


Here the result in a few sentences.

1.
I can't see any advantage of any OS over the other from the results
in this database.

2.
The AMD processors are about 20% faster with the "normal" solver.
So a 3800+ Athlon64 is about 20% faster as a 3800Mhz P4.
A 3000+ Athlon-XP is equivalent to a 3200+ Athlon-64.
The Intel-Centrino is also strong. Even in notebook, a
2GHz Centrino is equivalent to at least a 3.2GHz P4 desktop.
The P4 can reach the same speed as the AMD, if the compiler
is allowed to generate P4 optimized code. One time in the
past there was such a version available for LTspice. I remember
from Mike, that it's just too much effort to maintain two versions.
I have heared that the P4 and P3 is stronger with the
"alternate" solver compared to the AMD, but I haven't checked that.

The benchmark results are in the section "Database"
http://groups.yahoo.com/group/LTspice/



Back to the OS discussion.
If you experience a great difference with a certain simulation,
then the only reason for this can be the available cache memory
for the programm code and the data during this simulation.
Maybe this was the case when somebody thought it's faster
under Linux. I can tell you it's impossible in general, because the
processor runs exactly the same code, instruction by instruction,
regardless of the operating system. Only the I/O is different
between the OS, but this is only a small fraction of the simulation
task.

My statement:
There is absolutely no speed advantage to use LTspice under
WINE with Linux . I am conviced that a simulation can't be faster
there, when we look over many circuits in average. The good thing
is from what I see is that it's not slower. So you have the choice
with the OS.

Somebody in this group came up with the idea to use the processors
on the graphics card for the calulations. This is useless, because
you need 64bit floating point for SPICE. Otherwise a lot of
circuits will not run, because of the high dynamic range
and resolution required in the matrix solver.

Best regards,
Helmut

 

[qrk]

On Thu, 12 Jan 2006 14:46:07 +0000 (UTC), kensmith@green.rahul.net
(Ken Smith) wrote:

In article <1137066341.683300.280310@g49g2000cwa.googlegroups.com>,
Douglas Mota <douglasm@uninet.com.br> wrote:
Hi Ken,
Do you know why has LTSpice run faster over a 64 bit OS, even though
it's a 32 bit application? Sorry if it's a stupid question...

I think there are two reasons:

(1)
The 64 bit OS gets involved in the graphics operations and draws the
marching waveforms faster.

(2)
The spice engine writes a *.raw file. The 64 bit OS does the write
faster.


It could also be that some onrelated difference in the machines is the
cause. I have not run the two OSs on the same machine.

I hope this isn't a stupid answer

If your using an AMD 64-bit processor, the memory transfer rate is
!!way!! better than the old AMD Athlon CPUs and better than the newest
variants of the P4. Memory handling can account for a very large
difference in execution speed for some programs.

If your comparing different machines, beware that memory speed (not
the sticker speed, but actual measured memory transfer rates), size of
RAM, sustained hard drive writing speed, and background tasks (like
virus scanners) can have significant effects on program execution
time. Even BIOS version will affect memory transfer rates as seen on
my Tyan K8W motherboard. When comparing program execution times, you
need to benchmark various things, and then find out which component(s)
affect the execution time. It is not enough to say that the processor
or OS made up for all the time difference.

When comparing processors, you need to normalize the clock rate to
some number, like 1 GHz. Then you can see the relative efficiency of
the processes. AMD marketing came up with totally lame numbering
system that was designed to mislead the public. Use the actual clock
freq, not AMD's whacked out number scheme.

Interesting things I have found out about running a moderately complex
PSpice simulation on different machines:

1) AMD Athlon solved twice (assuming same CPU clock freq) as fast as
the old P4s, about 1.7x faster than the newer P4s. The PIII ran about
1.3x faster than the old P4.

2) AMD Athlon and Opteron are very close in execution time on Windoze
2000.

3) Memory speed had very little effect on execution speed on the
PSpice sim we were using.

4) Hard drive speed had a noticeable effect since a large data file is
written out.

My colleague benchmarked some Xilinx development tools and found that
things ran almost twice as fast on the Opteron system than his old
Athlon system, both using Win2k. This was mainly attributed to memory
speed. The Athlon memory transfer rates were less than half the rated
speed of the memory due to their really funky north bridge chip.

---
Mark

 

[Douglas Mota]

Somebody in this group came up with the idea to use the processors
on the graphics card for the calulations. This is useless, because
you need 64bit floating point for SPICE. Otherwise a lot of
circuits will not run, because of the high dynamic range
and resolution required in the matrix solver.

Hi Helmut,
-"...you need 64bit floating point for SPICE..."
Is it true even for 32bit Spice versions? Are there 64bit Spice
versions?
-"...a lot of circuits will not run, because of the high dynamic
range and resolution required in the matrix solver...."
Since I don't know nothing about this subject, there's my question:
so the processor on the graphics card does't have this required
processing power? Only the PC's CPU has it?

I'd really like to study and understand in depth the way Spice works,
including to learn enough to be able to modify the source code of some
of its modules. Could you please suggest me some material? It also
could be some web links and/or key words for searching in Google.

Thanks a lot!

 

[Ken Smith]

In article <1137504753.166437.220360@f14g2000cwb.googlegroups.com>,
Douglas Mota <douglasm@uninet.com.br> wrote:
[....]

-"...you need 64bit floating point for SPICE..."
Is it true even for 32bit Spice versions? Are there 64bit Spice
versions?

Beware that there are different meanings to the term "64 bit".

Most modern CPUs have a "bus width" that is commonly being refered to when
people say "16 bit", "32 bit" or "64 bit" when refering to a CPU chip.
The "bus width" in the number of bits the CPU can transfer in one action.
Basically, it is the number of wires going from point A to point B.

Like it is posible to do multidigit math "long hand" on paper, a CPU can
process numbers larger than the "bus width". You can do 64 bit math in a
8 bit processor this way. It won't be as fast but the results will be
exactly the same when it is done.

The x86 processors have a special section for dealing with the "floating
point" numbers. This section looks like it has an internal "bus width" of
80 bits. It may not really be 80 bits, since internally it could do two
40 bit transfers to get 80 bits from here to there.

While the numbers remain inside the "floating point unit" they remain in
the 80 bit form. When they are stored, they are rounded off to 32 or 64
bits as required.

-"...a lot of circuits will not run, because of the high dynamic
range and resolution required in the matrix solver...."
Since I don't know nothing about this subject, there's my question:
so the processor on the graphics card does't have this required
processing power? Only the PC's CPU has it?

You may have hit on something here. Just for purposes of discussion:
Assume that the graphics card only does 16 bit operations. Also, lets
assume that it does these operations at a 700GHz rate. Such a processor
would be able to outpace a 3.3GHz Pentium when doing 64 bit math.
Obviously, I just made up the 16bit at 700GHz values.

I'd really like to study and understand in depth the way Spice works,
including to learn enough to be able to modify the source code of some
of its modules. Could you please suggest me some material?

You can find the source code for some of the older versions of spice on
the web. The highly tweeked stuff that various companies provide is not
available. I assume by "modify the source code" you mean "modify the
source code in a useful way". To get to ths point the quickest, I suggest
you pick one part of the processing such as ".op" and dig into it in
detail.

It also
could be some web links and/or key words for searching in Google.

Spice 3f4 source

--
--
kensmith@rahul.net forging knowledge

 

[Stuart Brorson]

Douglas Mota <douglasm@uninet.com.br> wrote:
: I'd really like to study and understand in depth the way Spice works,
: including to learn enough to be able to modify the source code of some
: of its modules. Could you please suggest me some material? It also
: could be some web links and/or key words for searching in Google.

Here's the premier open-source SPICE. Linux native, with Mac and
Widows ports. Incorporates XSpice and Cider. Bugs which are present
in older SPICEs floating around the web are absent in ngspice since it
has been worked on continuously for several years.

http://ngspice.sourceforge.net/

Happy hacking!

Stuart

 

[Helmut Sennewald]

"Douglas Mota" <douglasm@uninet.com.br> schrieb im Newsbeitrag
news:1137504753.166437.220360@f14g2000cwb.googlegroups.com...

Somebody in this group came up with the idea to use the processors
on the graphics card for the calulations. This is useless, because
you need 64bit floating point for SPICE. Otherwise a lot of
circuits will not run, because of the high dynamic range
and resolution required in the matrix solver.


Hi Helmut,
-"...you need 64bit floating point for SPICE..."
Is it true even for 32bit Spice versions? Are there 64bit Spice
versions?

Hello Douglas,

This 64bit floating point math has nothing to do with the operating
system. It only menas that you use 64bits of memory for each variable
or constant. The calculations have to be also accurate for 64bit of
course. It's simply the type (double) in most programming languages.

-"...a lot of circuits will not run, because of the high dynamic
range and resolution required in the matrix solver...."
Since I don't know nothing about this subject, there's my question:
so the processor on the graphics card does't have this required
processing power? Only the PC's CPU has it?

I'd really like to study and understand in depth the way Spice works,
including to learn enough to be able to modify the source code of some
of its modules. Could you please suggest me some material? It also
could be some web links and/or key words for searching in Google.
Thanks a lot!

I have never tried to do that.
For me it's enough that Mike works in this area.
Insiders will know of whom I speak.

Best regards,
Helmut

 

[Stuart Brorson]

Stuart Brorson <sdb@cloud9.net> wrote:
: Douglas Mota <douglasm@uninet.com.br> wrote:
: : I'd really like to study and understand in depth the way Spice works,
: : including to learn enough to be able to modify the source code of some
: : of its modules. Could you please suggest me some material? It also
: : could be some web links and/or key words for searching in Google.

: Here's the premier open-source SPICE. Linux native, with Mac and
: Widows ports. Incorporates XSpice and Cider. Bugs which are present
: in older SPICEs floating around the web are absent in ngspice since it
: has been worked on continuously for several years.

: http://ngspice.sourceforge.net/

Also, if you are interested in analog circuit simulation in general,
you could take a look at Gnucap. Gnucap is a reasonably complete
open-source analog simulator which is a little more sophistocated than
SPICE. The code is also a lot cleaner, and so may be easier to
understand. Again, it's open source, so you can see and explore the
internals.

Here's the URL:

http://www.gnucap.org/

Rather than the 0.34 release, grab the latest snapshot available under
"development releases".

HTH,

Stuart