Standard die size (surface area)

[RoboTech]

Originally Posted by unregistered

I'd go 12 or 13mm sq

Why square? Why not rectangular (say 12x14)?
I doubt it would really make much difference, but rectangular would more closely simulate a CPU core, right?

[Groth]

Square is more challenging to cool (and we like a challenge).

Designers will lay out cores as square as possible to reduce the wafer area lost to cutting the dies apart.

[snowwie]

not all cores are rectangular, the prescott and northwood cores I think are square

also, not all cores are rectangular in the same direction, square is like a middle ground

[Huckleberry]

Now here's an idea: why don't they make the die hexogonal? They might be able to sqeeze a few more viable dies from the edges.

Of course parting the dies from the wafer would require a small miracle...

[Cathar]

Originally Posted by Huckleberry

Now here's an idea: why don't they make the die hexogonal? They might be able to sqeeze a few more viable dies from the edges.

Of course parting the dies from the wafer would require a small miracle...

Should be easily doable. Require 3 cuts instead of 2. Turning the wafer and aligning it to 60° boundaries would be a cake walk for the sorts of manufacturning technologies CPU makers are using.

Where it gets a little dubious is the amount of wafer material lost due to the extra cut between the dies, and whether or not that results in the same number of usable dies out of a single wafer at the end of the day. For larger dies, it would definitely give higher yields.

Would present some interesting layout issues trying to place down rectangular cache grids onto the hexagonal die though.

[BillA]

Originally Posted by Huckleberry

. . . . .

Of course parting the dies from the wafer would require a small miracle...

lol
ok Cathar, 3 scribed lines no problem
now how you gonna break it ?

[Groth]

This is a little later than planned, but...

Given 32 mm square copper heatspreader 1 mm thick, heat it with a 10 mm square 100 watt uniform heat source, cool the opposite side with a heatsink with an h(eff) of 20,000 W/m^2, what's the heat flux density at the heatspreader/heatsink interface?

[Cathar]

Originally Posted by unregistered

lol
ok Cathar, 3 scribed lines no problem
now how you gonna break it ?

Break it? I thought they were cut out using some sort of diamond-tipped cutting saw

[BillA]

noooo
scribe and break

Professor of Tea
how does the heat flux at the interface extend beyond the IHS ?

EDIT: how about an overlay ?
1) the CPU
2) TIM joint #1
3) IHS bottomside
4) IHS topside
5) TIM joint #2
6) sink underside

what you have shown would seem to be the sink underside, no ?

[Groth]

It doesn't. Look at the ticks and think about what the square in the center marks.

[BillA]

ah, . . .
well the 32 ticks I presume = the total IHS
so the 12 tick sq is the CPU

asleep at the wheel I guess

run another for the CPU/IHS interface

[RoboTech]

Very interesting Groth...

Am I understanding this correctly: the central red area has a heat density of 0.8 w/mm2, the area in orange is 0.7 W/mm2, and so on?

Would you do a model using actual CPU dimensions (P4 or Athlon 64) and making the concession of uniform heat ???

[Groth]

Middle square is only 10 by 10, naptime for Bill?

Uniform heat source on this one, CPU-IHS would be a very boring picture. Non-uniform sources are a whole 'nother can o' worms (when I put all 100W in one corner of the 'die', I got over 200°C corner to corner).

74% of the heat was transmitted through the 12x12 area indicated by the old rule of thumb of a 45° spreading angle.

[Groth]

Originally Posted by unregistered

how about an overlay ?
1) the CPU
2) TIM joint #1
3) IHS bottomside
4) IHS topside
5) TIM joint #2
6) sink underside

1) and 2) The CPU and TIM were modeled as a single unit for this one. More RAM and more experience first, then ?

3) The heat input was uniform, the flux map would be essentially a square step-function. Temperature map?

4) That what I've shown, heat flow perpendicular to the IHS top surface

5)I haven't put lateral heat conduction in my TIM, so the flux map will match the IHS top.

6)Again a matching flux map, due to no lateral in the TIM. Temperature map?

[Groth]

Originally Posted by RoboTech

Am I understanding this correctly: the central red area has a heat density of 0.8 w/mm2, the area in orange is 0.7 W/mm2, and so on?

Would you do a model using actual CPU dimensions (P4 or Athlon 64) and making the concession of uniform heat ???

The heat density at the underside is a uniform 1.0 W/mm^2 within the square shown. The map shows the heat density leaving the top of the spreader. Colors indicate a value between the values that bracket the color; thus red means between 0.7 and 0.8, and the black contour line 'tween red and orange would be 0.7 exactly.

Name the P4/A64 die size you wish.

[BillA]

measure it, 12 ticks (though perhaps you meant 10 ticks ?)

a uniform flux will not yield a uniform temperature across the die face when in contact with a larger sink, hence the flux will not be uniform either
my interest was the roll-off at the edges
really after the net difference between the bare CPU and the CPU + IHS

[Groth]

I don't have enough capacity/resolution yet for the roll-off at the CPU edges. If I add more nodes to the model, swap-file usage bring things to a crawl. Uniform die flux despite non-uniform die temperatures is a temporary kludge.

I can give an indicator of heatspreader effectiveness, but not anything finer yet.

Originally Posted by unregistered

measure it, 12 ticks (though perhaps you meant 10 ticks ?)

It's the weird spacing that's throwing you off -- I had to make the spacing between major ticks a factor of 32, to get things symmetrical. So a major tick is only four, not five. http://pages.sbcglobal.net/water.gro...cks-honest.gif

[RoboTech]

Originally Posted by Groth

The heat density at the underside is a uniform 1.0 W/mm^2 within the square shown. The map shows the heat density leaving the top of the spreader. Colors indicate a value between the values that bracket the color; thus red means between 0.7 and 0.8, and the black contour line 'tween red and orange would be 0.7 exactly.

Name the P4/A64 die size you wish.

Thanks, now I understand...

I can find the die surface areas (P4 = 146 mm2, P4 EE = 237 mm2, Athlon FX = 193 mm2) but not the actual dimensions of a die. If they were square a P4 would be ~12 x 12 mm and the Athlon FX ~14 x14 mm, but I don't know, so thanks anyway.

[Groth]

Just for you Bill, I simplified the waterblock to free up nodes and made a more realistic die. I also stole the RAM out of one of my folding boxes to give me a little more headroom.

The tea says ~36 hours for with the IHS, ~18 without. That is, if it runs (kinda pushin' my limits).

[BillA]

lol
I was chatting with a CFD guy and he asked what I would run it on, and when I told him of my big dual Athlon setup he could not stop laughing

bah

[Cathar]

Originally Posted by unregistered

lol
I was chatting with a CFD guy and he asked what I would run it on, and when I told him of my big dual Athlon setup he could not stop laughing

bah

Hmmm, seems a bit snobbish. Not even 10 years ago the place I worked at was hiring out one of the large 8-way Digital boxes that had 8 x 333MHz Alpha CPUs racked up with 8GB of memory, and that's exactly the kind of computational work that the system was used for. At the time, the system was one of the most powerful computers in the country.

A dual-Opteron with 4GB of ram would be perfectly capable of handling such problems, it's just that some modern engineers who laugh have forgotten where their roots were and the sorts of problems that were solved every-day on much weaker systems than are in use today.

For me as a software engineer, it's not my job to laugh at people, but to solve the problem. Despite the memory and computing shortfall next to a large computational cluster, the very nature of the problem is divide and conquer, except here you'd just be dividing amongst 2CPU's and 4GB of memory, instead of 128CPU's and 1TB of (distributed) memory. The computational problem is the same, it just takes longer, (or you just reduce the resolution).

Sorry Bill. Such attitudes cut at the core of my work ethos.

[BillA]

no Cathar, quite understand; but these guys deal in teraflops
the Athlons are 1200s on an early Tyan mobo
a modern desktop could cope if kinda maxed, over the weekend +

[Cathar]

Originally Posted by unregistered

the Athlons are 1200s on an early Tyan mobo
a modern desktop could cope if kinda maxed, over the weekend +

Ah. Computationally your system is not horrible. The amount of memory and the latency of the access to it is, more often than not, the main concern for these types of problems. Agreed, considerably slower if dealing with a dual-athlon 1200 box with ~1GB of (slowish) RAM. Definitely want to max out a modern high-end desktop platform if you want to pretend to limping along ("running" being too strong of a word) with the big boys.

Still, I wouldn't laugh at what can be accomplished with your system. Very similar to the sorts of systems we were dealing with a few years back in my present job, and when we got all the subsystems singing in harmony, it really isn't that much slower than a modern high-end system excepting in very specific tasks. Just takes a little more caressing.

[bigben2k]

Originally Posted by Groth

...

74% of the heat was transmitted through the 12x12 area indicated by the old rule of thumb of a 45° spreading angle.

Yikes!

Must check thumbs...

[BillA]

"caressing"
jesus christ you picked the wrong word there
last Sunday I got a blue screen, no boot, hdd failure

so being the IT co that we are, the hdd is now being doctored for several $k to recover a years worth of my keyboard pecking
shit
/rant