Sprinting CPU's?

[ehume]

"Computational sprinting"

Stuffing an i7 full of wax -- literally:

 

[KingFatty]

The Wired article didn't seem to mention any alternatives to using wax to absorb the momentary extra heat during the sprint.

But are there, or could there be developed, other materials like crystalline structures that absorb heat until they melt? Why is wax the best choice here to absorb lots of sprinting heat until it melts?

 

[ehume]

The aggregation page they cite is here.

 

[Greenlepricon]

The Wired article didn't seem to mention any alternatives to using wax to absorb the momentary extra heat during the sprint.

But are there, or could there be developed, other materials like crystalline structures that absorb heat until they melt? Why is wax the best choice here to absorb lots of sprinting heat until it melts?

I'm assuming because the wax is fairly cheap and it has a nice melting temp of 54C. It is a tad high imo for mobile devices, but it would seemingly do the trick. I'm by no means a materials scientist, but I can't think of anything else that is in the same price range with a good melting point. All I could find in the presentation linked was them referring to using a phase change material, and no specifics.

 

[Yuriman]

I'm skeptical when it comes to applying this to servers or desktops. I don't know as much about transistor design as many on this forum, but as I understand it, power use rises exponentially as you ramp up clock speed and voltage, so it's probably more efficient to, for example, spend 2 minutes at 2Ghz than to spend 1 minute at 3Ghz and 1 minute at 1Ghz. If the relationship were linear or even logarithmic, it would make sense. It's not like it's a race to put the computer to sleep and shut off the screen, which is generally the largest battery drain in mobile.

Additionally, they're suggesting that we design chips with a wider range of frequencies and voltages.

 

[Idontcare]

The Wired article didn't seem to mention any alternatives to using wax to absorb the momentary extra heat during the sprint.

But are there, or could there be developed, other materials like crystalline structures that absorb heat until they melt? Why is wax the best choice here to absorb lots of sprinting heat until it melts?

Yeah the point isn't to use a material which has a high heat capacity, but rather to use a material which has a high enthalpy of fusion.

High heat capacity just means the temperature keeps rising as you dump more heat into it, but the rate of increase in temperature will be low(er) because the heat capacity is higher.

Water is great in this regard.

But that doesn't keep your chip from over-heating and becoming a 90C furnace in your hand though.

To prevent the actual temperature from going up you need a material which is absorbing the heat while melting, because materials that melt do so at a fixed temperature.

Take ice for example, it doesn't matter how much heat or how fast you push the heat (within reason, no nuclear bombs obviously) into ice, the temperature will remain a fixed 0 Celsius until the ice has absorbed enough heat as to convert all the ice to liquid water, after which the temperature will then rise in proportion to its heat capacity.

Here is the heat absorbed vs. temperature for water:

A, C, and E are heat capacity - how much heat does it take to increase the temperature of ice, liquid water, and steam respectively?

B and D heats of fusion and vaporization respectively - how much heat does it take to melt ice and to boil water.

Notice how the temperature flatlines at those two phase transitions for the material despite all the heat being pumped into the system?

So the idea here is to have your sprinting microprocessor kept nice and cool while it is sprinting (lower leakage and more comfortable for a given form factor like being held by your hand) because you choose a material which has a low melting point and has a large enthalpy of fusion.

 

[zebrax2]

Wouldn't "sprinting" CPUs also need beefier parts on the power related side than those that are "jogging"?

 

[KingFatty]

Yeah the point isn't to use a material which has a high heat capacity, but rather to use a material which has a high enthalpy of fusion.

High heat capacity just means the temperature keeps rising as you dump more heat into it, but the rate of increase in temperature will be low(er) because the heat capacity is higher.

Water is great in this regard.

But that doesn't keep your chip from over-heating and becoming a 90C furnace in your hand though.

To prevent the actual temperature from going up you need a material which is absorbing the heat while melting, because materials that melt do so at a fixed temperature.

Take ice for example, it doesn't matter how much heat or how fast you push the heat (within reason, no nuclear bombs obviously) into ice, the temperature will remain a fixed 0 Celsius until the ice has absorbed enough heat as to convert all the ice to liquid water, after which the temperature will then rise in proportion to its heat capacity.

Here is the heat absorbed vs. temperature for water:

A, C, and E are heat capacity - how much heat does it take to increase the temperature of ice, liquid water, and steam respectively?

B and D heats of fusion and vaporization respectively - how much heat does it take to melt ice and to boil water.

Notice how the temperature flatlines at those two phase transitions for the material despite all the heat being pumped into the system?

So the idea here is to have your sprinting microprocessor kept nice and cool while it is sprinting (lower leakage and more comfortable for a given form factor like being held by your hand) because you choose a material which has a low melting point and has a large enthalpy of fusion.

Exactly, which is why I can't help but think a crystalline structure (like ice) would be excellent. Your wikipedia link to enthalpy of fusion suggests that wax is good, but crytalline structures are better? Sorry I'm not proficient in this subject matter, but it's very interesting.

 

[Homeles]

I'm assuming because the wax is fairly cheap and it has a nice melting temp of 54C. It is a tad high imo for mobile devices, but it would seemingly do the trick. I'm by no means a materials scientist, but I can't think of anything else that is in the same price range with a good melting point. All I could find in the presentation linked was them referring to using a phase change material, and no specifics.

The nice thing about paraffin is that it takes a lot of energy to convert it from a solid to a liquid. It'll hit its melting temperature, and stay there until enough energy is accumulated to turn it into a liquid.

In fact, I believe this is the primary reason for its usage in this application. It takes a lot of thermal energy for the wax to melt, which would mean that a CPU could put off quite a lot of thermal energy without getting cooked.

If I'm understanding this correctly, this should put things into perspective: water has a specific heat capacity of 4.183 J/gk; paraffin's heat of fusion is ~200-220 J/g. Basically, it takes 50x as much energy to melt paraffin than it takes to raise water 1 degree kelvin/Celcius at that same temperature.

So long as you aren't consuming too much power, the wax can absorb quite a lot of "thermal shock" as long as it is below its melting point, keeping the CPU nice and cool. However, wax is obviously not as good of a conductor as copper, so you have to maximize its surface area in order to get a good rate of transfer, otherwise the die won't be able to dissipate its heat fast enough and temps will run away. This is why the die has divots in it.

I may be oversimplifying things, may have bad math, or may have completely gotten everything wrong. I'm sure our resident chemistry nerd Idontcare will be in here to straighten things out (looks like he already poked his head in).

Exactly, which is why I can't help but think a crystalline structure (like ice) would be excellent. Your wikipedia link to enthalpy of fusion suggests that wax is good, but crytalline structures are better? Sorry I'm not proficient in this subject matter, but it's very interesting.

If I understand correctly, you need something that has a melting point as close to ambient as possible, while still being above it. Ice for example doesn't work because it is already liquid water at relevant temperatures. It also needs to have a high heat of fusion -- water has this, but both criteria need to be met. Silicon itself has a fantastically high heat of fusion at 1926 J/g, but because it doesn't melt until 1687K, it's not very useful for keeping temperatures around ambient

 

[Idontcare]

If I understand correctly, you need something that has a melting point as close to ambient as possible, while still being above it. Ice for example doesn't work because it is already liquid water at relevant temperatures. It also needs to have a high heat of fusion -- water has this, but both criteria need to be met. Silicon itself has a fantastically high heat of fusion at 1926 J/g, but because it doesn't melt until 1687K, it's not very useful for keeping temperatures around ambient

That is the crux of the reality of the implementation itself.

Crystalline or amorphous, that part isn't critical, what is critical is the ability to engineer the melting point and the enthalpy of fusion.

Amorphous materials made of polymers like wax are great for engineering the melting point, you simply choose how long you make the carbon chain and viola the melting point is dialed in.

Example

Engineering the melting point of a crystalline material can be a bit more challenging, not because the science or engineering is hard but rather because you need to be making a non-exotic material given that the consumer won't pay for anything high in price.

Wax is pretty darn cheap, so while not perfect it certainly gets the required job (low cost) done.

(^ I quoted you, Homeles, to build on your post, not to tell you stuff you already know)

 

[Saylick]

Ah yes... good ol' chemistry.

Two questions:

1) If all of the paraffin is melted (which would take a good amount of energy, no doubt), you're looking at 2.13 J/g-K of specific heat, roughly half the amount of water. What cools down the wax then? Would the heat dissipation of the wax fast enough to bring it back down to a solid state before the next sprinting cycle? Unless the whole idea is to never have all of the wax melt...

2) From what I understand, power (and therefore heat produced) scales exponentially with frequency. I understand that it can be better to "sprint" fast and park sooner than to "jog" and take longer to park, but can this philosophy really save energy if taken to the extreme as in the case of 100W smartphone processors? I figure that of the 100W, the last 30-40W or so only improves the clock rate of the chip by only 10-15% due to diminishing returns. Isn't this why Netburst was scrapped? It seems the whole point of this implementation is to be able to get snappier loading of web pages and whatnot but at the loss of power efficiency.

 

[Idontcare]

Ah yes... good ol' chemistry.

Two questions:

1) If all of the paraffin is melted (which would take a good amount of energy, no doubt), you're looking at 2.13 J/g-K of specific heat, roughly half the amount of water. What cools down the wax then? Would the heat dissipation of the wax fast enough to bring it back down to a solid state before the next sprinting cycle? Unless the whole idea is to never have all of the wax melt...

While it would not necessarily be a technical issue if you let the sprint cycle run long enough as to entirely melt the wax (or whatever phase-change material you are using), it is probably something that would be engineered to not happen.

The sprint time, or max power output as it were, would be intentionally limited such that the heat reservoir was never over-loaded.

The reason why I think they would ultimately implement it in that fashion is because you are essentially guaranteed to never exceed a specific operating temperature (the melting point) if you pursue it like that.

2) From what I understand, power (and therefore heat produced) scales exponentially with frequency. I understand that it can be better to "sprint" fast and park sooner than to "jog" and take longer to park, but can this philosophy really save energy if taken to the extreme as in the case of 100W smartphone processors? I figure that of the 100W, the last 30-40W or so only improves the clock rate of the chip by only 10-15% due to diminishing returns. Isn't this why Netburst was scrapped? It seems the whole point of this implementation is to be able to get snappier loading of web pages and whatnot but at the loss of power efficiency.

Power actually scales linearly with clockspeed.

The reason we observe, in practice, the power scaling exponentially with clockspeed is because leakage power scales exponentially with temperature. And as clockspeed goes up, pushing power up linearly, the temperature also increases, which results in the leakage power going up, which results in temperature going up, which pushes leakage up even more, etc.

It is a vicious upward cycle that stabilizes at a temperature which the heat dissipation machinery (the heatsink, fan, etc) and the ambient air temperature determine.

To further compound the problem, as temperature rises, so too must the voltage to ensure stable operation at a given clockspeed.

(signal to noise ratio, voltage is the signal, heat/temperature is the noise, higher temperature means more noise, so you must raise the voltage to keep the same signal:noise ratio or else the circuits fail to function at a given clockspeed)

This is why the traditional approach to sprinting (turbo clocks and the like) generally do not yield superior performance/watt results...because while "sprinting" the temperatures (and thus leakage power) do spike upwards in dramatic fashion.

But that is why they want to use a phase-change material, it caps the upper temperature, thus controllably preventing the leakage power from skyrocketing, which enables them to keep the voltage low as well.

It is brilliant, question is will it be economical once engineered to be robust enough to deal with the daily abuses our mobile form-factor devices suffer?

 

[KingFatty]

HAhaha now I'm imagining desktop applications, where instead of seeing who has the biggest heatsink/watercooling loop setup, people start comparing who has the biggest block of wax, with some making elaborate sandcastle types designs out of their wax units.

What I was thinking is that the wax temperature will not go over the melting point temperature as long as you aren't fully melted. But once all the wax is melted, the temp will rise and the chip can detect the sudden rise in temps and declare that uptick as the end of the sprint period.

So if you get a huge wax block, you get more sprinting time because it sucks down more heat before it melts. Practically speaking, I bet there is a physical limit where the effect doesn't work because part of the block touching the CPU will entirely melt but a far-away section would not melt and would actually be thermally insulated from the melting part by all the wax. Oooh, maybe they can infuse the wax with diamond dust or other particles to give the wax great thermal heat transfer properties like a heatsink, in addition to it's ability to guzzle down heat while melting?

 

[BUnit1701]

... Oooh, maybe they can infuse the wax with diamond dust or other particles to give the wax great thermal heat transfer properties like a heatsink, in addition to it's ability to guzzle down heat while melting?

I presume that is the purpose of the 'metal mesh' described in the article.

 

[Idontcare]

Yeah you still need to manage the thermal conductivity itself, otherwise you will end up with a sprinting CPU with a skyrocketing temperature because the heat itself can't get to the heatsink (the wax).

On a desktop that would probably look something like taking a stock fan and inserting it into upside down into a ball of wax (think of making a candle, only use the HSF instead of the candle wick).

Then take that ball of wax and insert it into an even larger HSF, the wax fitting inside an enclosed internal cavity to dissipate heat to the ambient.