new spice

>"John Larkin"  wrote in message 
>news:bdc7lgdap8o66j8m92ullph1nojbg9c5ni@4ax.com...

>https://www.linkedin.com/in/mike-engelhardt-a788a822

...but why ......?


-- Kevin Aylward

http://www.anasoft.co.uk/ SuperSpice
http://www.kevinaylward.co.uk/ee/index.html 

On Thu, 30 Sep 2021 19:58:50 +0100, "Kevin Aylward"
<kevinRemoveandReplaceATkevinaylward.co.uk> wrote:

>>"John Larkin"  wrote in message 
>>news:bdc7lgdap8o66j8m92ullph1nojbg9c5ni@4ax.com...
>
>>https://www.linkedin.com/in/mike-engelhardt-a788a822
>
>...but why ......?
>

Maybe he enjoys it. I'm sure he's enormously wealthy and could do
anything he wahts.

I want a Spice that uses an nvidia board to speed it up 1000:1.

-- 

If a man will begin with certainties, he shall end with doubts, 
but if he will be content to begin with doubts he shall end in certainties.
Francis Bacon

Am 30.09.21 um 21:24 schrieb John Larkin:
> On Thu, 30 Sep 2021 19:58:50 +0100, "Kevin Aylward"
> <kevinRemoveandReplaceATkevinaylward.co.uk> wrote:
> 
>>> "John Larkin"  wrote in message
>>> news:bdc7lgdap8o66j8m92ullph1nojbg9c5ni@4ax.com...
>>
>>> https://www.linkedin.com/in/mike-engelhardt-a788a822
>>
>> ...but why ......?
>>
> 
> Maybe he enjoys it. I'm sure he's enormously wealthy and could do
> anything he wahts.
> 
> I want a Spice that uses an nvidia board to speed it up 1000:1.
> 

Hopeless. That has already been tried in IBM AT times with these
Weitek coprocessors and NS 32032 processor boards; it never lived
up to the expectations.

That's no wonder. When in time domain integration the next result
depends on the current value and maybe a few in the past, you cannot
compute more future timesteps in parallel. Maybe some speculative
versions in parallel and then selecting the best. But that is no
work for 1000 processors.

The inversion of the nodal matrix might use some improvement since
it is NP complete, like almost everything that is interesting.
Its size grows with the number of nodes and the matrix is sparse since
most nodes have no interaction. Dividing the circuit into subcircuits,
solving these separately and combining the results could provide
a speedup, for problems with many nodes. That would be a MAJOR change.

Spice has not made much progress since Berkeley is no longer involved.
Some people make some local improvements and when they lose interest
after 15 years their improvements die. There is no one to integrate
that stuff in one open official version. Maybe NGspice comes closest.

Keysight ADS has on option to run it on a bunch of workstations but
that helps probably most for electromagnetics which has not much
in common with spice.

Cheers, Gerhard

On 9/30/2021 5:12 PM, Gerhard Hoffmann wrote:
>> I want a Spice that uses an nvidia board to speed it up 1000:1.
> 
> Hopeless. That has already been tried in IBM AT times with these
> Weitek coprocessors and NS 32032 processor boards; it never lived
> up to the expectations.
> 
> That's no wonder. When in time domain integration the next result
> depends on the current value and maybe a few in the past, you cannot
> compute more future timesteps in parallel. Maybe some speculative
> versions in parallel and then selecting the best. But that is no
> work for 1000 processors.

I think a factor of two *overall* is possible with a bit of work.  I've
seen proposals that suggest a factor of 4 is likely the upper limit on
any such work.  Some models may hit sweet spots -- but not the sort of
thing you're going to boast about in sales literature!

These are small enough that it's easier just to twiddle your thumbs and
wait for next year's PC update cycle to get that improvement "for free"
(i.e., for no development work cuz new work means new bugs).

> The inversion of the nodal matrix might use some improvement since
> it is NP complete, like almost everything that is interesting.
> Its size grows with the number of nodes and the matrix is sparse since
> most nodes have no interaction. Dividing the circuit into subcircuits,
> solving these separately and combining the results could provide
> a speedup, for problems with many nodes. That would be a MAJOR change.
> 
> Spice has not made much progress since Berkeley is no longer involved.
> Some people make some local improvements and when they lose interest
> after 15 years their improvements die. There is no one to integrate
> that stuff in one open official version. Maybe NGspice comes closest.

A big part of the problem is folks want these tools for free.
That attitude suggests they place no value on the time they save
running simulations (or, that their time *has* no value).

Someone will come up with a breakthrough and sell it as a *service*.
So, you'll submit your circuit and get a bill for the analysis.  Of
course, this will likely cut down on the "what if" use of such a
tool (cuz changing a portion of the circuit will still require a complete
REanalysis... likely hard to leverage any prior computations to reduce
the cost of those efforts)

Then again, software-as-a-service is the obvious future.

> Keysight ADS has on option to run it on a bunch of workstations but
> that helps probably most for electromagnetics which has not much
> in common with spice.

Am 01.10.21 um 05:47 schrieb Don Y:
> On 9/30/2021 5:12 PM, Gerhard Hoffmann wrote:
>>> I want a Spice that uses an nvidia board to speed it up 1000:1.
>>
>> Hopeless. That has already been tried in IBM AT times with these
>> Weitek coprocessors and NS 32032 processor boards; it never lived
>> up to the expectations.
>>
>> That's no wonder. When in time domain integration the next result
>> depends on the current value and maybe a few in the past, you cannot
>> compute more future timesteps in parallel. Maybe some speculative
>> versions in parallel and then selecting the best. But that is no
>> work for 1000 processors.
> 
> I think a factor of two *overall* is possible with a bit of work.  I've
> seen proposals that suggest a factor of 4 is likely the upper limit on
> any such work.  Some models may hit sweet spots -- but not the sort of
> thing you're going to boast about in sales literature!

You are probably right here.

> These are small enough that it's easier just to twiddle your thumbs and
> wait for next year's PC update cycle to get that improvement "for free"
> (i.e., for no development work cuz new work means new bugs).
> 
>> The inversion of the nodal matrix might use some improvement since
>> it is NP complete, like almost everything that is interesting.
>> Its size grows with the number of nodes and the matrix is sparse since
>> most nodes have no interaction. Dividing the circuit into subcircuits,
>> solving these separately and combining the results could provide
>> a speedup, for problems with many nodes. That would be a MAJOR change.
>>
>> Spice has not made much progress since Berkeley is no longer involved.
>> Some people make some local improvements and when they lose interest
>> after 15 years their improvements die. There is no one to integrate
>> that stuff in one open official version. Maybe NGspice comes closest.
> 
> A big part of the problem is folks want these tools for free.
> That attitude suggests they place no value on the time they save
> running simulations (or, that their time *has* no value).

No. The big part of the problem is that spice 2g4 or 3.x was always free
and some people added some miniscule things like a GUI  and expected
they would rule the world from now on since THEY had written THE
simulation program. For a time that worked for few of them, like pspice
that got swallowed by Orcad that got swallowed by Cadunz. And where
are they now? When have you seen the last result from pspice and its 
offspring?

JT from s.e.d. is RIP now for some years, but do you think he missed any
of the new developments? No, not really. Which new developments?



> Someone will come up with a breakthrough and sell it as a *service*.
> So, you'll submit your circuit and get a bill for the analysis.  Of
> course, this will likely cut down on the "what if" use of such a
> tool (cuz changing a portion of the circuit will still require a complete
> REanalysis... likely hard to leverage any prior computations to reduce
> the cost of those efforts)
> 
> Then again, software-as-a-service is the obvious future.

You mean, like H-Spice some decades ago?
OMG, I'm growing old.
Some customers of mine don't like it when I have their stuff on MY
computer, much less in a cloud.

Cheers, Gerhard

On 9/30/2021 9:32 PM, Gerhard Hoffmann wrote:
> Am 01.10.21 um 05:47 schrieb Don Y:
>> On 9/30/2021 5:12 PM, Gerhard Hoffmann wrote:
>>>> I want a Spice that uses an nvidia board to speed it up 1000:1.
>>>
>>> Hopeless. That has already been tried in IBM AT times with these
>>> Weitek coprocessors and NS 32032 processor boards; it never lived
>>> up to the expectations.
>>>
>>> That's no wonder. When in time domain integration the next result
>>> depends on the current value and maybe a few in the past, you cannot
>>> compute more future timesteps in parallel. Maybe some speculative
>>> versions in parallel and then selecting the best. But that is no
>>> work for 1000 processors.
>>
>> I think a factor of two *overall* is possible with a bit of work.  I've
>> seen proposals that suggest a factor of 4 is likely the upper limit on
>> any such work.  Some models may hit sweet spots -- but not the sort of
>> thing you're going to boast about in sales literature!
> 
> You are probably right here.

Technological speedups are a red herring.  What you're really concerned with
is the TOTAL time to perform a particular action.  If you speed up some
portion of it 100-fold... but, that was just 20% of the entire process,
what have your *real* gains been?

There's also a "sour spot" (if something is the opposite of "sweet spot"
I would imagine "sour spot" to be it?) where things get faster... but still
not fast *enough*.

E.g., decades ago, I would render 3D models and it would eat 100% of a
machine's time for 36 hours.  It would be *silly* to sit and wait for
the result!  I'd simply power off the monitor, put a sign on the
machine reminding me NOT to turn it off, then move on to some other
task.  Some time later (36, 40, 72 hours??), I'd come back to the
machine and, hopefully, find the result that I desired.  But, in
the mean time, I'd have worked on something *else*, making
uninterrupted progress, there.

If rendering took 10 minutes, I'd likely have twiddled my thumbs for
those 10 minutes -- too little time to do much of anything beyond
get a drink or bite to eat.  So, I've spent 10 additional minutes of
"personal time" on the work (that wasn't present in the original
approach).

Worse, if the cost of waiting gets small enough, you change your
work strategy from one of thoughtful preparation to "hit-and-miss"
("Let's try this and see how it works?").  So, you end up performing
more iterations than you might, otherwise.

[I know many engineers who are perpetually tweaking designs
instead of DESIGNING them from the outset!]

>> These are small enough that it's easier just to twiddle your thumbs and
>> wait for next year's PC update cycle to get that improvement "for free"
>> (i.e., for no development work cuz new work means new bugs).
>>
>>> The inversion of the nodal matrix might use some improvement since
>>> it is NP complete, like almost everything that is interesting.
>>> Its size grows with the number of nodes and the matrix is sparse since
>>> most nodes have no interaction. Dividing the circuit into subcircuits,
>>> solving these separately and combining the results could provide
>>> a speedup, for problems with many nodes. That would be a MAJOR change.
>>>
>>> Spice has not made much progress since Berkeley is no longer involved.
>>> Some people make some local improvements and when they lose interest
>>> after 15 years their improvements die. There is no one to integrate
>>> that stuff in one open official version. Maybe NGspice comes closest.
>>
>> A big part of the problem is folks want these tools for free.
>> That attitude suggests they place no value on the time they save
>> running simulations (or, that their time *has* no value).
> 
> No. The big part of the problem is that spice 2g4 or 3.x was always free
> and some people added some miniscule things like a GUI  and expected
> they would rule the world from now on since THEY had written THE
> simulation program. For a time that worked for few of them, like pspice
> that got swallowed by Orcad that got swallowed by Cadunz. And where
> are they now? When have you seen the last result from pspice and its offspring?

So, they were "spoiled" by a boon -- and now are expecting that
"good will" to continue?

What would you *pay* for 1000X speedup in your simulations?

Would you offer up the value of the (your!) time saved over,
say, a calendar year?

Or, would you grumble at what you perceived as "price gouging"?

People have funny ideas of what items are worth -- when they
have to open their own wallets!  :>

> JT from s.e.d. is RIP now for some years, but do you think he missed any
> of the new developments? No, not really. Which new developments?

Some of that can be cultural.  I write all my code on a *BSD box running
X -- with a "root weave" wallpaper.  No naked ladies, no pictures of
fields of flowers, no tweets and bops and whistles.  Doesn't change
the quality of the code I produce to be absent those things.

OTOH, I use some costly tools to help me test my codebase as they
are more reliable than "other eyes".  And, I can "ask" them to
test over and over again without "overstaying my welcome".

>> Someone will come up with a breakthrough and sell it as a *service*.
>> So, you'll submit your circuit and get a bill for the analysis.  Of
>> course, this will likely cut down on the "what if" use of such a
>> tool (cuz changing a portion of the circuit will still require a complete
>> REanalysis... likely hard to leverage any prior computations to reduce
>> the cost of those efforts)
>>
>> Then again, software-as-a-service is the obvious future.
> 
> You mean, like H-Spice some decades ago?

You have to offer genuine value, not just an alternate user
interface.

Autodesk offers SfM as a service -- they sell you a GUI that
essentially just acts as an input terminal for your data and
a presentation portal for their results.  You can do roughly
the same thing on your own workstation -- but with many many MIPS
(and some dubious results).  Some tools being "free" -- but
often with blemishes that you won't discover until you've *invested*
(!) some time -- for which you  may not see any return!

> OMG, I'm growing old.

Heh heh heh... a common malady!

> Some customers of mine don't like it when I have their stuff on MY
> computer, much less in a cloud.

That's one reason all of my machines (save this one) are air-gapped.
There's no time wasted reassuring clients (or myself) that I've
not been hacked!  And, no *effort* spent keeping my defenses "current"!

OTOH, bean counters see services as a cost saving.  Why have your own
IT department to perpetually update and service bought (ahem, "licensed")
tools when you can have someone else sort all that out, for you?!

I've noticed (time tracking) that I tend to spend about a day a week
on "support" issues -- keeping equipment running, looking for
upgrades, ditto software tools, backups, chasing down bugs/workarounds,
etc.  It's easy to see how outsourcing those things (or parts of
them) can be attractive.

[OTOH, *not* outsourcing means I have control over how and when
those tools change/evolve.  I can opt to "live with" some known
problem instead of having it replaced by a new set of UNknown
problems without my consent]

Part of the appeal of web-based services is that all the user
(client) needs to do is keep a current browser running.  Even
an IT department made of dweebs can likely do this (esp if
they run diskless clients)

On 2021-10-01 05:47, Don Y wrote:
[...]
>
> Then again, software-as-a-service is the obvious future.

I'll be kicking and screaming the whole way.

Jeroen Belleman

On 9/30/2021 11:03 PM, Jeroen Belleman wrote:
> On 2021-10-01 05:47, Don Y wrote:
> [...]
>>
>> Then again, software-as-a-service is the obvious future.
> 
> I'll be kicking and screaming the whole way.

Just say "no"!

What you have *today*, works.  Keep that in mind for
tomorrow... and the day after... and...

(Do you really *need* version N+m of whatever tool is working
for you *now*?)

Am 01.10.21 um 07:56 schrieb Don Y:

> Technological speedups are a red herring.  What you're really concerned 
> with
> is the TOTAL time to perform a particular action.  If you speed up some
> portion of it 100-fold... but, that was just 20% of the entire process,
> what have your *real* gains been?

Amdahl's law.

<   https://en.wikipedia.org/wiki/Amdahl%27s_law  >


Cheers, Gerhard

On 10/1/2021 2:15 AM, Gerhard Hoffmann wrote:
> Am 01.10.21 um 07:56 schrieb Don Y:
> 
>> Technological speedups are a red herring.  What you're really concerned with
>> is the TOTAL time to perform a particular action.  If you speed up some
>> portion of it 100-fold... but, that was just 20% of the entire process,
>> what have your *real* gains been?
> 
> Amdahl's law.
> 
> <   https://en.wikipedia.org/wiki/Amdahl%27s_law  >

Misses the point.

There's still an organic being *in* the process.  More cores
doesn't make that being more productive in *its* actions,
thought processes, etc.

E.g., if you have to *think* about how you want to change
(or create) a design between iterations, that's a cost
that a faster processor won't help reduce.  Likewise with
actually editing the design.

(unless you just POKE at the design until you see results
that you like -- because you don't really understand how
it actually works and are relying on the simulation to
give you a "feel" for that)