heat die dimensions - a public worklog

[BillA]

the intent of this worklog is to develop a dimensioned drawing for a heat die reflecting present industrial practice
- while there is nothing to prevent a DIYer from making and using such a device, the required resources may be substantial
- while this is a specific activity of myself/CoolingWorks, it is hoped that a general consensus might result in an open design available to anyone wishing to utilize such
(Ben jump in here - but not your WBTA please)

the general configuration of a heat die is well understood as can be seen in this old Aavid paper
http://www.aavidthermalloy.com/techn.../packaging.pdf
and there are recent papers in the last several years by both Intel and IBM describing the same configuration
-> and most notably is the 'amazing' work to the same point by 'our' Incoherent (3 cheers !)

there is nothing magical about heat dies, different designs properly executed can yield valid and comparative results as this paper by Sun illustrates
http://www.electronics-cooling.com/h...05_feb_a2.html

as most here know, a heat die is 'also' a TIM joint tester and it is well to be aware of industry practice, progress, and goals;
such can be seen here http://www.electronics-cooling.com/h...vember_a2.html
and a TIM joint appraisal here http://www.electronics-cooling.com/h...vember_a1.html
and here http://www.electronics-cooling.com/h...bruary_a1.html

thermal test dies and dimensions of contact areas are reviewed here
http://www.coolingzone.com/Content/L...Sep2000_3.html

the above list is not complete, but sufficient to understand that our efforts are not in a vacuum; we do not need to reinvent the wheel (and lack the resources too, eh ?)

I have a question not addressed in the above papers:
Assuming a calorimeter type sensor placement, what is the necessity to use copper for the heat die ? (i.e. does the matl's conductivity matter so much ?)
while Ben's suggestion to use iron seems radical, why not use a more durable material ?

It seems apparent that the size/configuration of the testing heat source must be related to the actual source size; and the wear of the copper face introduces an known and unacceptable variation in the test results (limited initial degradation progressing at an ever increasing rate).

regarding the face dimensions, I'm inclined to start with 2 sizes, 10x10mm and 12x12mm - based more on current desktop CPUs than any trend

and for the heat source 2 heater cartridges in the base

comments solicited

[jaydee]

Glad you started this thread.

I am working on a drawing of what I was going to build based heavily on Incoherents work. Will post when finished.

As for the material to use the main reason I went with copper is because the block material is usually copper. If you use something harder the chances of damaging the base of the water block increases with multiple remounts. Steel would be a very very bad thing to use IMO. Copper, aluminum or brass were the only materials I was going to use. Stuck with copper for conductivity reasons, maybe not relevant though?

The main problem I have with heater cartridges is finding a low ohm version. I would prefer them at 5-10 ohm but very hard to find them under 100ohm...


Anyway hope this thread takes off.

[BillA]

await your dwg

for years HSFs were put on silicon, don't see hardness per se as limiting - has very much to do with the mounting procedure and force applied
- but a sharp corner and a tilted mount will cause damage; to the die if made in copper and to the HSF if made in steel
-> as the die is vastly more expensive, I would opt for protecting the die

eventually you will understand that you need a big linear psu

[jaydee]

Here is some pics of my first drawing (and only drawing so far). I went with 1/2" die just for convenience and can easily be changed. Note the holes for the probes are .25" deep. The 4 holes on the outside sides are meant to be tapped for screws to keep the cartridges tight. I was planning on using 1" long and 1/4" wide 100watt 120V cartridges available at mcmaster.

What other materials would you suggest we look at?

[BillA]

re the dwgs:
- set screws on bottom to push up
- use 1" stock for more thickness on bottom (if not copper ?)
- decrease length to 1.25" if 1" cartridges are used
- use two 200W AC heaters

what was/is the basis for your sensor to sensor to face dimensions ?
noting that the 2ed sensor to face dimension will change over time (reworking) if of copper

someone have experience with calorimeters ?
what is the trade-off with 2 sensors vs. 3 ? (2 points a line, 3 a curve - no ?)
Incoherent, have you looked at this ? (did I miss a previous discussion ?)

a material ?
Silicon Carbide, SiC ?? ( some info http://www.accuratus.com/silicar.html)

[jaydee]

Originally Posted by unregistered

re the dwgs:
- set screws on bottom to push up
- use 1" stock for more thickness on bottom (if not copper ?)
- decrease length to 1.25" if 1" cartridges are used
- use two 200W AC heaters

what was/is the basis for your sensor to sensor to face dimensions ?
noting that the 2ed sensor to face dimension will change over time (reworking) if of copper

someone have experience with calorimeters ?
what is the trade-off with 2 sensors vs. 3 ? (2 points a line, 3 a curve - no ?)
Incoherent, have you looked at this ? (did I miss a previous discussion ?)

a material ?
Silicon Carbide, SiC ?? ( some info http://www.accuratus.com/silicar.html)

- Set screws! That's the term I was looking for. No problem on changing that to bottom.

- I was assumining we wanted as little material as possible but if not I can easily add more to the bottom. Would benifit the set screws aswell.

- It is 1.25" no?

- If 2 - 200 watts heaters are recommended I would have to go with 1.5" long.

Sensor holes were based on ease of drilling more than anything else. To close to the top or bottom would be hard to drill.


Not sure on the rest. Will change things as suggested.

Silicon Carbide looks pretty good. Were can it be bought and how easy is it to machine?

[BillA]

er, strictly diamond tooling I believe
SiC are cutters, no ?

way big $ for a chunk I'd guess, Monday I'll source some
I have phenolic BTW if you need a piece

note that what we are doing will 'work' with any material

[RoboTech]

I still find brass interesting (my first thermal die sim was brass, but later versions have all been copper)

+Thermal conductivity similar to silicon
+Tougher than copper (more robust surface area)
+Easy to machine, readily available
+Relatively good thermal conductivity
- Main body will run much hotter for a given die surface temp than copper, especially when incorporating a calorimeter tower section. This will produce much greater heat loss to ambient. Can be easily accounted for though by running a heat loss calibration curve.

[BillA]

good suggestion Lee
silicon bronze is tough stuff too
I'll dig deeper

yea re the energy balance
previously I super insulated but at the cost of huge time constants, 4+hrs to equilibrium
better to accept and simply quantify the secondary losses

[bobo5195]

Ive seen thermal rigs which instead of insulation used heaters

Ie heatsource - some insultation - heater.

Heater temp is aligned to current heatsource. No temp differential, no heat transfer. Equilbrum should be fast but added complexity

I'll have a look around for stuff. I know some people in my year did a precise thermal test rig for plastics last year but i should have a paper with ideas somewhere or other.

[BillA]

B5
yea, a guard heater
be a bit of a buggar in an environmental chamber - but doable

an interesting matl is cu-co-be UNS C17500 with a thermal conductivity of 180 and a Rockwell B hardness of 100 vs. 37 for cu
- I cannot find a hard brass with any conductivity, and bronze is worse

the combination of thermal conductivity and hardness probably has many answers, as do the material plus machining cost implications

[Cathar]

Originally Posted by unregistered

the combination of thermal conductivity and hardness probably has many answers

Industrial diamond?

[BillA]

"as do the material plus machining cost implications"

mercy, I don't know if I can handle carbide !

[jaydee]

Modified drawing pics attached. Left out sensor holes for no idea where to put them.

Changes?

[bigben2k]

Oh so much to cover (I go away for one day, and so much activity!!!). I'll tackle in varying order.

As for the die material: I have to consider what's commonly available, so I refer to our sponsor:
http://www.onlinemetals.com/index.cfm?affiliate_id=302
(link from Main page)

- Invar (an alloy of 36% nickel and 64% iron)
- Nickel Alloy R405 (similar to Monel)
- Titanium
- Tool steel (including "ALLOY 1018")
- Stainless

Maybe dubious purity, would not care to see 392, 396, 400 debated again.

My concern however is with the maintenance. Was thinking about some type of plating followed by a grind to restore die to same total height. The idea still needs thought, but all in all, copper's softness is a problem.

Damage to the heat die could be avoided/minimized with a proper mounting plate that would have "guides", but without interfering (significantly) with the final alignment. The damage expected would be a rounding of the contact area's edges.


As for dimensions, if we "stretch" (!) the Jedec 85% uniformity requirement to 85% of actual die size, then 12 by 12 mm will not do (extended, 12 by 12 is useless, see below).

I've gathered many actual core dimensions (a hard exercise!) but found most of the info at geek.com . For Intel, excluding the old Willamette core, core sizes vary between 112 mm2 and 237 mm2. That means using two dies, 10 by 10mm and 14 by 14mm, which actually does cover every core, within 15%. I have not collected all data for AMD.

Below is the collected data, complete with source links:

"CPU core dims and power"

Intel:

Smithfield 90nm
206 mm2 / 95 to 130 Watts
from THG

Prescott 2M 90nm
135 mm2 / 84 to 115W
from: http://www.digital-daily.com/cpu/intel-prescott-2m/

Gallatin 130nm
237 mm2 / 54 to 97 W
from: http://www.geek.com/procspec/intel/p7server_13_4.htm

Prescott 90nm
112 mm2
from: http://www.extremetech.com/article2/...1478683,00.asp
103W
from: http://techreport.com/reviews/2004q1...t/index.x?pg=2

Northwood 130nm
131 mm2 / 61 to 68W
from: http://www.geek.com/procspec/intel/northwood.htm

Willamette 180nm
217 mm2
from: http://www.geek.com/procspec/intel/p7consumer.htm
52 to 72W
from: http://www.cpuscorecard.com/cpuprices/ip4.htm

AMD:

Toledo/Manchester 90nm

San Diego 90nm
115 mm2 / 104 W
from: http://www.tomshardware.com/cpu/2005...n_fx57-02.html

Clawhammer 130nm

Venice 90nm

Newcastle 130nm

Winchester 90nm

Sledgehammer 130nm

Clawhammer 130nm

Clawhammer old 130nm

Newcastle 130nm

Paris 130nm

Palermo 90nm

Thoroughbred B 130nm

Thoroughbred A 130nm

Barton 130nm

Thorton 130nm

[BillA]

10x10 and 14x14
ok by me, 14 should serve for GPUs moderately well ?

[bigben2k]

Oumpf. the 15% margin isn't there (after further review).

Let me gather the AMD data first.

[bigben2k]

Some more info:
http://www.tomshardware.com/cpu/2004...istory_big.gif

I'll recompile.

[jaydee]

Originally Posted by bigben2k

Some more info:
http://www.tomshardware.com/cpu/2004...istory_big.gif

I'll recompile.

Link dose nothing.

[Les]

Copy and paste works
God, it is a big list

[jaydee]

Originally Posted by Les

Copy and paste works
God, it is a big list

Had to close all browsers first then paste and it worked....

Big list indeed. Any idea on tomorrows die sizes?

[bigben2k]

Originally Posted by Les

Copy and paste works
God, it is a big list

...and quite the lovely puzzle too:
-Clawhammer on 754 and 939
-Barton on Athlon XP and Sempron
-I may have skipped over the mobile and servers...

I think that this exercise is just going to be about completing the missing elements of the THG chart (below):

(updated as needed)

Toledo/Manchester 90nm
199 mm2 / 110 W

San Diego 90nm
115 mm2 / 104 W

Clawhammer 130nm (socket 939) model 4000+
144 mm2 / 89W
from: http://geek.com/procspec/amd/k8.htm

Clawhammer 130nm (socket 939) model 53 / 55
144 mm2 / 89W

Venice 90nm
84 mm2 / 67 to 89W
http://www.xbitlabs.com/articles/cpu...-venice_3.html
(core size "same as Winchester").

Newcastle 130nm
144 mm2 / 89W

Winchester 90nm
84 mm2 / 67W

Sledgehammer 130nm
193 mm2 / 84.7 W

Clawhammer 130nm (socket 754)
193 mm2 / 89 W

Clawhammer old 130nm (socket 754)
193 mm2 / 89 W

Newcastle 130nm
144 mm2 / 89 W

Paris 130nm
118 mm2 / 89 W

Palermo 90nm
> 84 mm² / 62 W
http://www.presence-pc.com/tests/Ath...Palermo-295/3/
(in french, but states "> 84 mm²", closest available estimate. Consistent with various other statements)

Thoroughbred B 130nm Athlon XP model 1700+ to 2800+
84 mm2 / 59.8 to 74.3

Thoroughbred B 130nm Sempron model 2800+ to 3?200+
101 mm2 / 59 W

Thoroughbred A 130nm
80 mm2 / 49.4 to 67.9 W

Barton 130nm Athlon XP 2500+ to 3000+
101 mm2 / 68.3 to 76.8 W

Barton 130nm Sempron 3000+
101 mm2 / 62 W

Thorton 130nm
101 mm2 / 60.3 to 68.3 W
http://forums.amd.com/index.php?showtopic=53302
("Basically, it's a 'Barton' with half of the L2-cache disabled.")

[jaydee]

Also are we going to forget the IHS exists or make a compensation?

[BillA]

the die
JEDEC
IHS is well known, been discussed ad nauseum; go to the lit

[Ice Czar]

Originally Posted by unregistered

but sufficient to understand that our efforts are not in a vacuum

or are as the case may be

Thermal Contact Resistance Measurments under Vacuum @ U of Waterloo 2003
via Stanford S.T.E.P.

pretty much your first link

Experimental Apparatus & Proceedure
All measurements were performed using the thermal interface material test apparatus at the MHTL shown in figure 1. This apparatus, as described in the publication included in the Appendix of this report uses a pair of calibrated heat flux meters with equally spaced RTD elements to measure the total heat flow rate through the joint. Heating is provided by 4 cartridge heaters in a copper block at the bottom of the test column, a liquid cooled cold plate at the top of the column acts as a heatsink for the system. Loading is perfornmed using a linear actuator connected to a lever system, and a 1000lb load cell is used to measure contact pressure at the joint. The entire measurement apparatus is contained in a vacuum chamber, and all tests are performed under vacuum conditions, p<5Pa (0.037torr). A Keithley 2700 data aquisition system is used to perform all the measurements, and data logging and control of the experiment are performed using Labview v.5.i software running on a Windows-based PC.

Two approaches are traditionally used to stabilize or eliminate heat losses, i) a guarded heater where surrounding conditions are controlled through a secondary heater and ii) a vacuum environment where conduction and convection heat losses are minimized and radiation heat losses can be controlled through a radiation heat shield.

the construction notes in the appendix are quite detailed
and relating to the actual question they used Aluminum 2024 for the flux meters

it looks like because of the tight dimensional control of the flux meters they are able to measure the temperature gradient and with the thermal conductivity of the Al 2024 determine the heat flow rate.
larger size of the flux blocks to get that gradient being the reason I assume they opted for a vacuum over the added complexity of insulation and secondary heaters when employing the RTDs
so provided the same data is available for other alloys and they are truely homogenous...