Storm/G7 Prototype + Piccies

[bobo5195]

Originally Posted by Les

Complete crap.
Focus on Cathar's postulations and apply any privileged(University Library Passwords) to constructive/destructive analysis

Wha?

[Les]

Not sure I understand "Wha?"
However Cathar's predictions are clear:-
With a ~100mm² CPU die size, and 6W of hydraulic pumping power, my waterblocks have exhibited roughly the following historical progression of effective convectional transfer efficiencies:
"Concept Block" - original WW precursor with plastic slit nozzle Prototype: ~62000 W/m²K
White Water x 1mm channels: ~67000 W/m²K
White Water x 0.8mm channels Prototype: ~71000 W/m²K
Free (un-nozzled) Jet Array Against flat base-plate (1mm jets) Prototype: ~55000 W/m²K
Mini-cupped free jet (un-nozzled) Prototype: ~65000 W/m²K
Cascade: ~72000 W/m²K
Cascade SS: ~74000 W/m²K
Cascade XXX Prototype: ~80000 W/m²K
Storm/G1 Prototype: ~50000 W/m²K
Storm/G3 Prototype: ~65000 W/m²K
Storm/G4: ~77000 W/m²K
Storm/G5: ~85000 W/m²K
Swiftech STORM (G4 Rev2): ~83000 W/m²K
Storm/G5 w/ G4 Rev2 optimisations (theoretical): ~91000 W/m²K (projected estimate)
Storm/G5 w/ G7-level optimisations (theoretical): ~94000 W/m²K (projected estimate)
Storm/G7: ~105000 W/m²K

Are these proposed h(eff) values correct ?

[bobo5195]

i was talking of

(real c/w)/ideal(c/w) = efficantcy

where real c/w is pefromance minus tim and possibly a thin base plate of material which is needed to prevent excessive transfer.

Ideal c/w = dt/P = 1/(density * heatcapacity * flow)

Or even usign the differential of the produced graph

[bobo5195]

excel crashed on me but i was getting values a straight line graph with values of a where y=ax+b of about 8 to 10 (efficantcy/lpm)

[Les]

Originally Posted by bobo5195

i was talking of
(real c/w)/ideal(c/w) = efficantcy

A new definition.
So what.
The proposition was Cathar's proposed values for h(eff)
My analysis has been presented(link
Is it correct?

[bobo5195]

think we've crossed wires i was talking about representing the preformance of blocks and how much of a boost there is left to go. i prefer my method for the simple reason that it gives a nice straight line as opposed to a 1/x style relationship. Although as a graph i will accept that without reprocessing into w/k style its not helpful.

My method is essentially the stanton number (st=h/(rho * V * cp)) which is another form of the nusselt number, of the maybe 20 papers ive read they all use the nusselt or stanton (althought they call it nusselt) number instead of c/w.

I cant see any holes and i in no way disagree with the way its presented genrally. Im not so sure about non dimensionalising area when talking about computers (this is a nitpick, in the case of your data non dimensionalised area is correct), in the case of test rig it is quite helpful, but for a cpu die where the cpu heating area is smaller than the heat spreader or is what is actually used. If block A cools more area than block B it shouldnt be penalised for that if the same cpu is being used. A different cpu die size is a different kettle of fish as its bigger size does not result in linear increase in performance for a same geometery block. Im talking different blocks here of different designs, for the case of the G seris 1/m^2 corelation is useful.

edit: spelling

[Cathar]

Originally Posted by Les

A new definition.
So what.

Well, it's the same definition, just worded differently.

Originally Posted by bob5195

(real c/w)/ideal(c/w) = efficantcy

where real c/w is pefromance minus tim and possibly a thin base plate of material which is needed to prevent excessive transfer.

The watts are constant, so they cancel out.

We're left with efficiency = Real C / Ideal C.

As bobo says, this is all minus tim and bp, aka, the surface temperature where the convection(conduction) is taking place. Since this is relative to the water's temp, it's the exact same measurement of efficiency that you referred to above Les, and that I was talking about.

I haven't had the time to dig further into Les's analysis of h(eff) yet. Will do so this weekend.

As for nusselt & stanton numbers, I'll leave that stuff to the theorists to quantify. h(eff) is pretty much all I'm concerned about as a means of assessing design performance.

[bobo5195]

Originally Posted by Cathar

Well, it's the same definition, just worded differently.



The watts are constant, so they cancel out.

We're left with efficiency = Real C / Ideal C.

As bobo says, this is all minus tim and bp, aka, the surface temperature where the convection(conduction) is taking place. Since this is relative to the water's temp, it's the exact same measurement of efficiency that you referred to above Les, and that I was talking about.

I haven't had the time to dig further into Les's analysis of h(eff) yet. Will do so this weekend.

As for nusselt & stanton numbers, I'll leave that stuff to the theorists to quantify. h(eff) is pretty much all I'm concerned about as a means of assessing design performance.

Well its (c/w) * Q as we are only concerned with water. Cp and rho drop out and the units are only constants * lpm for example. Considering how wb perform according to c/w = k/Q its more useful than c/w for a specific flow rate as a mesure of performance.

reading my heat transfer textbook its actually amazing how useful the stanton number is. Its effective non dimensional heat flux/ non dimentsional heat transfer coefficant. You could design a fairly good water block of just the stanton number and its member parts. st= nu/(re*pr)

[Cathar]

In a system in equilibrium, Watts is constant too, however the Watts is used to derive the water temperature rise for whatever mass flow rate is occurring.

[Cathar]

Originally Posted by bobo5195

You could design a fairly good water block of just the stanton number and its member parts. st= nu/(re*pr)

No, you could quantify how effective a design is using those numbers.

A dimensionless number alone means little, other than as a means to quantify some effect. Can't design a waterblock around a dimensionless number. Can use the number to know that you're headed in the right direction, but h(eff) is suitable enough for that purpose anyway, especially in a compound scenario of many effects.

[bobo5195]

Surely your first though is the non dimensional numbers as they give you all the important parameters. Alright you mihgt need an intial "good" block to work from but it should help.

the modified reynolds analogue/chilton colburn analogies state that the averge boundary layer thickness is

St*Pr^2/3 for hydrodynamic condesiderations

and

St.mean*Sc^2/3 for themral boundary layer
[sc = schmidt number, ratio of momentum and masss difusivities. Cc=kinematic viscosity/D)

From that you can guess a good width of channel. Fine equations are known so that can be approximated well to start off with as well. Fancy stuff like jet impingement cant be done first off, but jet impingement numbers are alwasy given dimensionless and you should be able to scale results using the St number.

That tell yous you how big to make the fins on a block so that a good amount of fluid is going to get through. For a specific gemometry it allows you to scale sizes up and down.

[Long Haired Git]

Nope.
'Been drinking since 2pm due to a mate's last-day and a big deal win, and I was hoping it would help, but sober or drunk, none of this makes sense.

[mwolfman]

SUPER OFFTOPIC
EDIT: Page down, showed a quad nvidia 7800 SLi system and the next P4, power supply 1500Watt... try to cool that with air in a case...

[Bloody_Sorcerer]

congrats, that is the most off-topic post I have ever laid eyes on.

[Bundles]

Pretty much what i thought when i saw it too.

[mwolfman]

Don't say that I dident warn you! (even told what it showed!)

[bobo5195]

yes it is most off topic congrats. With 4 gpus you can be processor limited twice as much as normal. Shucks with a single 7800gtx your processor limited on just about every game at most proper res'.

Anyway to say something on topic. Ive just a had a read of a paper comparing fluid properties with heat transfer. Its not good news for fans of fluids other than water. Water dominates by quite abit. They were trying to compare prandlt numbers of various fluids but waters properties are so above that of others it completely skews the result. For fluids (as opposed to gases) they found that water had a local heat trransfer coeff (w/m^2 K of about 900,000 while other fluids had about 300,000.

The test was carried out on fleunt, which is a very good cfd package however i still have doubts about the usefulness of cfd with jets, especially without specifying a particular boundary layer equations (there are about 6 common ways of calculating boundary layers in cfd, some are better than others for jets). They were also using low Re of maximum 100 (i believe the g7 is something like 600) and a wide nozzle dia of 5mm. An isothermal plate is used which is good for modeling but is not much help for computers because with such high heat transfer coeff biot numbers are going to be high (greater than one, aka convective heat transfer is greater than the conductiveness of copper) so the copper cant get heat to the jet center fast enough. The reason they postulate for waters goodness is its high heat transfer coef. They reckon that this is the governing variable for heat transfer to a fluid in jet impingement. This is basedd on their studies of gases.

They commented on two more effects. A spike in heat transfer at x/dia (where dia is jet diameter) at 20, this is caused by a local lowering in heat transfer at about 19 then it suddenly shoots up a little (30% maybe). My theory is this could be due to turbulent lamina transition. It only happens in high prandlt number fluids (liquids basicly). They also observe that heat transfer increases if you bring the jet close to the surface, due to a more confined faster jet (as postulated by other papers, i think theres other stuff going on as well). Other papers have observed this. However this effect is noted to increase pressure drop faster than c/w so super close might not be benefical. Increased surface heat transfer under the jet might not be optimal due to the biot relation (copper cant get heat to the water fast enough) so spreading out the jet which lowers its max c/w but increases the area which high c/w under the jet occurs. It should be noted that for an unconfined jet moving the jet closer (and so the restricting wall) always increases heat transfer as the fluid is moving faster, however far away from the jet the amount of heat transfer may not be useful.

Anyway hope your all still with me. And long haired git your best with another beer :P. Sir Issac newton has just given me a way to calculate fluid flow in hideously complicated piping systems with very little itteration.

[mwolfman]

I hope Il get some time next week (still working on a friday, 00:42) then Il try to do a CFD on the intire block, thats my way of doing things... Then we will se that the heat transfer is and where and if there are any part of the block thats unessesery (pressure/mass loss)
Might do a rouff one to start with and then zoom in on the center jet...
We are talking about a 10^2mm die (or?) but what load? 130 W or is it an OC cpu?

[Cathar]

Interesting reading all of bobo's speculations. These are all a subset of the considerations that went into the design of the Storm over time.

bobo, I do have to ask. Where are you going with this? Are you concerned that perhaps I was not aware of such considerations, or are you just out to slowly work your way towards understanding why the Storm is designed the way it is, which is fine.

I guess I'm trying to figure out how much involvement you want from me in this, or if you're just happy to keep throwing out the research and ideas which were all considered when designing the block. I think that what you'll find if you continue on for long enough is that the design is a very complete and well balanced marriage of an even broader range of concepts and effects than what you're commenting on.

I suppose I'm just saying that I'm somewhat enjoying watching you take the path to discovery about the design, and asking myself if I should assist or just watch. Seems to me that you're on the trek to design the block from first principles, which is fun to watch I have to admit.

[GlassMan]

A lot of the original thinking is here G4
http://forums.overclockers.com.au/sh...d.php?t=299417. there is probably an equivelent thread at this site as well.
Earlier thinking here Cascade
http://forums.overclockers.com.au/sh...hreadid=169230

[Cathar]

Originally Posted by GlassMan

A lot of the original thinking is here G4
http://forums.overclockers.com.au/sh...d.php?t=299417. there is probably an equivelent thread at this site as well.
Earlier thinking here Cascade
http://forums.overclockers.com.au/sh...hreadid=169230

Yeah, that's some of the more lower-level explanation stuff. What bobo's doing above is starting to touch on the stuff I did. I wasn't too concerned about calculating the Nusselt or Stanton numbers though, as I stated, such are just theoretical observations of a working system. Problem I had was that the cooling effect was a lot more complex and not readily modeled by existing research, so aside from taking a cue from existing research to give a good guideline and to help establish some good educated guesses at where to start with various physical parameters of the design, beyond that it more came down to a combination of theory and empirical data, and once those steps had been exhausted the last few 0.1C's were gained almost purely by empirical playing.

The existing theory will get you ¾ of the way there, but beyond that it pretty much comes down to directed experimentation.

[bobo5195]

theres alot ot theory out there but its light on parameterisation. There are many people who have been designing coolers like this for decades but they are not going to tell you all their secerts. The models very close to being complete, from waht ive read theres enough out there that alot of needs to be said has been done but its an exceptional difficult problem. The use of turbulators is very rarely mentioned for example and it is quite obvious that some commerical systems use this but they might be little understood outside the intuition of the designer.

Models of jets hitting plates are used as a vary good way of testing CFD systems they are a problem case that its very easy to get wrong. Not alot has be done to test confined jets for the simple reason that they are what you used if heat transfer is what you are after and nobody wants to get off the gravy train. One paper mention 3 orders of magnitude higher C/W for a confined jet.

The paper i mentioned above does show one thing that nobodies mentioned. If heat transfer coeff of the metal is the governing factor of the fluid then liquid metal cooling should be very very good with jet impingement. This depends on how viscous the metal is of course and if it can ever be used like that with little pumps.

The second point it does make is that local c/w under the jets are very high. Higher than metal conductivities. I cant quite get my head around it if this means that upgrading metal conductivity will give a disproportionate boost in perfromance as heat can get to the cool zones better and that base plate thickness should reflect the need to move alot of heat to some points. Or that metal thermal conductivity is notimportant as the jets are cooling very well by themselves.

My main goal of this is a good cfd simulation where you just click a design and wait for abit as there is no other way to establish the heat disribution easily for designs like this. Also there is abit of designy stuff into the bobo flow approximator if i ever get around to making it ( i really java and one of my FEA packages uses python so i think i might try and write some stuff in that). It would be nice to know what a good approximation for PQ and C/w vs Q for that to save me using a long forumla and assuming that test rigs are "good".

[mwolfman]

What heat load is the block intended for and for what die size (the °C/W isn’t a constant factor, it depends on the heat load, IE: we should revaluate the °C/W-factor and include the die area and heat load somehow in the factor)?

(My f*king computer support has shut down the license server... (its not responding anyway) so I can’t do anything until I get my hands on the idiot, I don’t have the key to the server room...)

[Les]

Originally Posted by bobo5195

.......... It would be nice to know what a good approximation for PQ and C/w vs Q for that to save me using a long forumla and assuming that test rigs are "good".

It would be nice, but this is the problem
There is not " a good approximation for C/w vs Q", if you are referring to measured "C/W".
Good data for "Swittech Storm"(G4) but as dicussed here interpretation is impossible without more details.
Good data for LRWW but requires leap of faith re C/W(TIM)(different TIM and wb surface finish) and possibly "sensor offset allowance" (Die conductivity(360-400w/mk and sensor position(seperately reported as 1.9939mm))
Possibly the only "good and interpretable" data is the SwiftechMCW6000 on a 144sq mm heat-source (Incoherent's presentation here and subsequent posts).

[bobo5195]

sorry i may of said this wrong.

I mean a proper non dimensional scaling. Pipes/ channel blocks are v^2. Thats how they vary. You can run as many tests as you want but unless theres some odd reason due to geometery or something thats what they run like. For jet impingement style situations this approximation may be correct but i would prefer a nice good quality paper to tell me so.

That does not in anyway mean that the given results are wrong or a bad approximation at all. It just meas i can be completely confident in them, espeically as they often dont have error bars and the such like or independent verfication.

All the data ive seen on reputable sites with good method is perfectly acceptable (well my docs would give even the best a C, maybe D :P but its still good enough, which is i more often than not got ).