WC simulator discussion.

[deeppow]

Originally posted by gmat
I'd be glad if someone provided me with an algorithm for a Bernstein one-dimensional method (also called B-Spline by some ppl, though i thought it was Bezier)

I don't have one but if you do a search on "Bezier-Bernstein approximation" you'll find a number of links. Maybe that'll help.

[gmat]

I've been quite out of free time lately. Looking for a written algorithm (in any language, i'll translate) is not that easy.
What i'm looking for is:
- given a set of points in the Euclidian plane: Pn={x;y}
- find a suitable function y=f(x) so that if i input {Pn;x} the function return an interpolated y (and y is a real).
That means, a 1D continuous function.
I'm looking for the Bernstein method for this, as it seems to produce more precise polynoms than Lagrange method, and which are less prone to the 'wavelet syndrome'.

[Alchemy]

Originally posted by gmat
That's what i'm doing right now, a small app that will allow anyone to put a XML file and see the interpolated results on a graphical chart.
Currently i have implemented two polynomial interpolation methods, that seem to produce very similar polynoms, with differences in the order of 1.10^-16. I think the choice will be made on performance as the algorithms involved are totally different (Lagrange and Neville).

Not sure what you mean by such tiny differences.

What are the R-squared values, and how many points are you interpolating from?

Alchemy

[deeppow]

Originally posted by gmat
I'm looking for the Bernstein method for this, as it seems to produce more precise polynoms than Lagrange method, and which are less prone to the 'wavelet syndrome'.

gmat, you just passed me. All I remember bout the Bernstein polys is that they're slow to converge, I think. Nor am I sure what the wavelet syndrome is. Not sure an answer belongs in this thread, maybe too much of a diversion.

[arhines]

Has anyone thought of using something designed to do numerical modeling? Matlab?
I'm sure some of you have read the thermal simulation article at ocers.com (cant find the link in their nebulous archives, but I assure you it's there). In any case, matlab is very adaptable to all sorts of simulations, and is extremely equation friendly. I bet someone around here has registered versions of some thermal simulation toolboxes they can use at least at work. This would be a valuable tool to consider - very powerful, very easy to implement.

[gmat]

Originally posted by Alchemy
Not sure what you mean by such tiny differences.

What are the R-squared values, and how many points are you interpolating from?

Alchemy

which root squares ? The polynomial interpolation i use produce a function such as every point supplied in the dataset is a control point, and such as the resulting polynom passes exactly by control points.

The tiny diffenrences are produced at interpolated values, between 2 different methods. Given the order of those differences i consider them as zero.
I haven't tested complex / weird datasets though, so i don't know how which method fares at limits.

The 'wavelet syndrome' is oversensivity of interpolated data on one end of the set when you 'move' one point on the other end of the data set. The resulting function overcompensates and produces sort of 'waves' on one end. This is easily corrected by a more complete data set but it's worrying nevertheless.
I'll provide a small test tool so you can see by yourself.

(matlab): it's not a programming language, and it does nothing a real programming language can't do. Besides the sim is meant to be distributable to anyone as an executable program (and sources), and free. Matlab doesn't answer these needs.

[Skulemate]

Originally posted by arhines
Has anyone thought of using something designed to do numerical modeling? Matlab?

That would cause problems if those involved in this simulator wanted to distribute it... and it would prevent ProCooling from hosting a version on their site. Course, that assumes that open availability is one of the aspects of this collaborative effort.

[Since87]

I'd like to see this eventually implemented as a Java applet that is hosted by ProCooling.

Having a variety of revisions of the application, and un-validated component models floating around could cause unwarranted credibility problems. Better IMO to have a central location where somewhat expert review of proposed models is done.

[Skulemate]

I like that idea, though I also like the idea of having a core program that can deal with separate input files that each represent a component in the system... that way you could add components as Bill tests them, and making revisions to the core would only require a quick redownload.

[Since87]

Originally posted by Skulemate
I like that idea, though I also like the idea of having a core program that can deal with separate input files that each represent a component in the system... that way you could add components as Bill tests them, and making revisions to the core would only require a quick redownload.

Ahh! I see an incentive program. Those who contribute 10 good models to the central database get a copy of the core program.

They can be among the first to run a simulation with the latest thing Bill's tested.

[BillA]

hmmm
shall we cut to the chase ?

I do hope a bunch of you have some serious stroke with the mfgrs to get them to come up with data for their products
(or have it done 'for' them)

perhaps this 'calculator' would be an incentive in and of itself ?

[Since87]

I'm hopeful that the calculator could be popular enough to result in pressure on mfgs to submit products for testing.

My impression from observing the chatter at Overclockers, is that your Waterblock Bench Testing Results article has greatly reduced the frequency with which Maze3's are recommended. My best guess is that a (decently lapped) Maze3 is very similar in performance to a TC-4, but without hard data, few will stand up for it.

I think the availability of the calculator will greatly improve understanding of the relevance of the data, and will result in sales going to products where people can gauge performance before purchase.

Of course, a certain critical mass of accurate models will be necessary before the simulator will become popular enough to have an impact. Aside from rads, I think enough 'modellable' data has been published to reach critical mass.

FWIW, I sent off email to Hardware Labs, suggesting they send their products to you for testing. Others wanting to do the same can email them here:

info@hwlabs.com

What "chase" shall we "cut too"?