Why Hspice somtimes get wrong results when using multicore?

[Arthur Asimov Heinlein]

I use -mt command to make hspice run on several cores, it sometimes gets correct results faster than just one core running, but sometimes gets totally wrong results while I could get correct results by running on just one core. Why's that case?

Thanks!

 

[ahmad1954]

maybe that version of hspice is not optimized for multicore processing.

 

[Arthur Asimov Heinlein]

maybe that version of hspice is not optimized for multicore processing.

I am using the latest version hspice 2014.... Could it be the problem of precision?

 

[dick_freebird]

If your circuit has multiple valid solutions (as many do) then
accuracy and method choices, coupled with the vagaries of
numerical solution goal-seeking, can give you different and
yet valid solutions; error residues can be different and then
differently amplified, etc.

Multicore is all abotu solution speed, and speed-first often
means looser accuracy defaults and more error residue at
every solution-point. You might try enforcing total consistency
of solution method, steps and tolerances and see what results.

You might want to look closer at the "wrong" solution and
how you got there, because it may just be "unexpected"
and waiting to emerge in your design validation sample
in real life.

 

[Arthur Asimov Heinlein]

If your circuit has multiple valid solutions (as many do) then
accuracy and method choices, coupled with the vagaries of
numerical solution goal-seeking, can give you different and
yet valid solutions; error residues can be different and then
differently amplified, etc.

Multicore is all abotu solution speed, and speed-first often
means looser accuracy defaults and more error residue at
every solution-point. You might try enforcing total consistency
of solution method, steps and tolerances and see what results.

You might want to look closer at the "wrong" solution and
how you got there, because it may just be "unexpected"
and waiting to emerge in your design validation sample
in real life.

Thanks for your answer, but could you please tell me how to enforce total consistency of solution method, steps and tolerances?

 

[dick_freebird]

I don't use HSPICE but the SPICEs I do use, can be made to
dump the variables list; your tolerances and method are
controlled by some of them. Their names, you'd probably
have to go to the user docs for (I could tell you what I see,
but it might not help).

The steps, ought to just be evident in the .tran statement.
However you might want to read up on what happens in the
two different modes when, say, you don't specify the minimum
timestep; different processing might assert different defaults,
perhaps.