The Old Joel on Software Forum: Part 3 (of 5)

high k gate dielectric

There was a recent discussion on microsoft switching to the PPC due to heat troubles with the intel line.

Intel's made the breakthrough in physics/materials science needed to solve that problem:

http://www.intel.com/pressroom/archive/releases/20031105tech.htm?iid=HPAGE+low_news_031105b&

Summary: When this process goes on line in 2007, we'll have absurdly high clock rates with almost no heat dissapation and extremely low power requirements. Fans and liquid coolers will be a thing of the past. Etc, etc. Beers all round.

Dennis Atkins
Tuesday, November 11, 2003

> Summary: When this process goes on line in 2007, we'll have absurdly > high clock rates with almost no heat dissapation and extremely low >ower requirements. Fans and liquid coolers will be a thing of the past. >Etc, etc. Beers all round"

No, heat disaapation and power requirement will always be a matter of concern. It certainly makes both of these less of an issue to designers however. The biggest thing it will bring to the table is allowing Moore's Law (transistor density) to continue. A 700 million transistor microprocessor running at say, 8GHZ, running on a supply voltage of <1v, on 45nm tech, could very well have a TDP of 40-90 Watts.

Jim Battin
Tuesday, November 11, 2003

Thermodynamics works regardless of your marketing budget.

Simon Lucy
Wednesday, November 12, 2003

Totally offtopic with this discussion, however the thermodynamics comment caught my attention as it's something that I see on the message boards quite frequently. It generally appears in the context of a discussion of the sort "today we have A that uses 1X energy to accomplish Z, with 0.9X waste, but tomorrow we'll have B that uses 0.15X energy to accomplish Z, with 0.05Y waste" (be it cars, computers, etc). Almost without fail there is a stock "yeah, but what about the laws of thermodynamics?" (on Slashdot they usually get moderated up as the voice of cynical wisdom). These comments generally tend to be of the "obey the status quo" variety, and are presented under the belief that anything that is more efficient than what we have as the norm today must somehow break the laws of thermodynamics. This is astoundingly untrue.

Dennis Forbes
Wednesday, November 12, 2003

Well the same variables and their ratios hold true if you use the same methods. Apply a brake made of material X and a wheel hub of material Y and you can change the amount of wear with the amount of material but you won't affect the basic equation. You'd need to change one or other of the materials involved to make any difference, or find some other way of slowing the wheel.

And so each time processor speed and inefficient heat dissipation have looked like a brick wall to increasing transistor densities a new wrinkle in dissipating that heat has been found.

But if you don't change something radical you're stuck with the same ratio.

Simon Lucy
Wednesday, November 12, 2003

Oh, I spose I'd better clarify even that.

Voltage leakage is one input into the generation of heat but its not the only one, I don't even think its the most important one. So reducing voltage leakage is no doubt a good thing and it will allow greater densities of transistors but if the clock speed is also increased then heat is going to be much the same still.

Simon Lucy
Wednesday, November 12, 2003

ACM Queue theme last month was power management and affects on Moore's law.

http://www.acmqueue.org/modules.php?name=Content&pa=showpage&pid=76

One article (not online) has a graph suggesting chip power consumption could reach that of a nuclear reactor sometime in the next decade or two!

MugsGame
Wednesday, November 12, 2003

Simon, I hate to do it since you're usually a rasonable chap, but you're kind of confused about what you are talking about here.

#1 - current leakage, not voltage leakage is the problem.

#2 - your comments on the conservation of energy are not relevant because the amount of heat energy created by a transistor is anything BUT a constant. For the current family of processors, energy is only lost during switching. There is virtually no leakage when the transistor is in one state. Higher clock rates mean more switching and thus more leakage. Lower processor voltages help by reducing the leakage during the switch but are more sensitive to noise. Increased densities do increase heat of course. But there is nothing to stop physicists from inventing a transistor that does not leak current during switching, which is the case for transistors using this new dielectric material with a higher k constant.

Dennis Atkins
Wednesday, November 12, 2003

s/does not leak/does not leak as much

Dennis Atkins
Wednesday, November 12, 2003

The gate leakage is a static drain of current (DC). While it must be significant - or else intel would not care, its not the only source of heat generation.

The higher the clock rate the more the dynamic (AC) current is the source of heat (wow, is this an oversimplification, and my apologies to any intel or other solid state engineers out there). Anyhow, capacitance (which is a function of gate geometry and dielectric premetivity - and as mentioned in the article, the gate is ony 5 atoms thick!) in part, governs the top end clock frequency. The more capacitance provided by the gate, the more charge is required by the preceding transistor to turn off/on the channel.

Bottom line: even if they completely solve static drain, the capacitance of the gate will still be a problem that inhibits clock frequency - so it will take smaller and smaller gate widths, channels and transistor spacing to maintain higher clocks.

Was this a more important problem for Intel to solve than say a PowerPC fabricator? Does CISC (x86) require a higher transistor count than PowerPC equivalents? (Are there any PowerPC equivs?) I don't know... my knowledge of such things is so dated its embarassing.

hoser
Wednesday, November 12, 2003

hoser:

Maybe, but perhaps becoming less of an issue. I honestly have no idea how many transistors are used in a modern x86 architecture to translate instructions, but considering the relative success of the platform I wouldn't consider it terribly significant. As transistor densities increase, the amount required to handle the x86 instruction set becomes less significant when you look at the big picture.

Jim Battin
Friday, November 14, 2003