I was reading an article about Radeon 6990 recently and it was recommending 1200 Watts of power supply and the entire machine to be kept dipped in mineral oil because air cooling would kill it in a few months if not. My question is why are chips converting so much electricity into heat? What does physics say into this and is it not possible using the very advanced technologies we have today to make our chips a bit more efficient?

Edit1: I am hoping for a technical answer. Apologies if that was not evident in the original question.

  • 1
    \$\begingroup\$ I reformatted the title of your question a little bit. First, IC stands for integrated circuit, not chip. Second, not all ICs get hot, so I added "some" to the title. \$\endgroup\$ – Connor Wolf Jun 6 '13 at 20:06

Why does any machine get hot when heavily used? Because nothing is 100% efficient. Wires have resistance. Transistors have gate capacitance. Every flow of electrical energy necessarily has some inefficiency. Energy can't be destroyed, so where does this lost electrical energy go? Heat. When you further try to cram as much functionality and speed into the smallest space possible (GPUs being at the leading edge of this endeavor), you get a lot of heat in a small space.

Believe me, engineers are using the most advanced technologies they can to reduce the heat generated by GPUs. The main thing that prevents them from making them faster is that they would burst into flames if they did. Every day, they find ways to make the processors more efficient, so they can cram in more functionality and more speed to take them back to the threshold of self-destruction by heat.

See also Why is the CPU hot when it's turned on?

  • 2
    \$\begingroup\$ I think the main root cause of heating is gate capacitance; billions of tiny little caps being charged and discharged billions of times a second. Almost as effective as shorting the supply. Also why shrinking feature size is so attractive, as below. \$\endgroup\$ – Nick T Jun 6 '13 at 18:10
  • 1
    \$\begingroup\$ Historically, when a technology for reducing heat generated is discovered, we compensate by turning up the clock speed until the chip is back at the same temperature. The performance vs. power game is changing, though, because so many applications are battery-powered. \$\endgroup\$ – markrages Jun 6 '13 at 18:12
  • \$\begingroup\$ Interesting comment there @NickT. If its the capacitors and they are almost 'shorting' each time, then heat generation is a fundamental issue that seems it like cannot be attacked directly. Thx for the added info. \$\endgroup\$ – Regmi Jun 6 '13 at 18:40
  • \$\begingroup\$ @PhilFrost - Thanks for the answer but I was hoping to get a more technically detailed answer. Perhaps it was the way I asked my question that it was not evident...(added through an edit.) \$\endgroup\$ – Regmi Jun 6 '13 at 18:44
  • \$\begingroup\$ @Regmi the technical answer is in there: wires have resistance, and transistors have gate capacitance. If you don't understand those concepts, I'd suggest asking a new question about them. \$\endgroup\$ – Phil Frost Jun 6 '13 at 19:34

Yes, it's possible to make ICs more efficient, at the expense of speed. The problem is that a one is created by charging the gate capacitance of a mosfet to a certain value. If you want to change that bit to a zero, you must discharge that capacitor.

The determining equation for power in a CMOS process for active power is:

$$ P_{active}=CV^2f*A$$

C = capacitance, V = Voltage, f = frequency(clock speed) and A = activity factor.

You add up billions of gates and the heat becomes significant. A theoretically 100% efficient machine would have infinite propagation delay, at least until someone discovers a better way than using voltage to signify a bit. Until then, the only ways to reduce heat are:

  • Reduce clock speed.
  • Reduce gate capacitance, usually by making smaller transistors.
  • Reduce voltage.
  • Turn off parts of the chip you don't need.

This is the approach taken in battery-powered applications. I have a 1 GHz ARM board that runs completely cool to the touch without fans or heat sinks. That's a remarkable improvement over a 1 GHz desktop IC from 10 years ago or so, but the current bleeding edge computers take those same improvements and crank up the speed, so it stays just as hot.

  • 1
    \$\begingroup\$ The major problem with your curve is that when we have MOSfets driving MOSfets they look like current sources driving the capacitive gates which is actually a straight line not an exponential like you show. It only gets curved as you near the triode and sub-threshold regions of operation of the transistors. \$\endgroup\$ – placeholder Jun 6 '13 at 18:47
  • \$\begingroup\$ Feel free to edit. My VLSI is a little rusty, which is where my "more or less" came from. The exact shape of the curve doesn't affect the rest of the answer, though. \$\endgroup\$ – Karl Bielefeldt Jun 6 '13 at 18:55
  • \$\begingroup\$ -1, there isn't an inherent link from speed to efficiency. As Mark pointed out in a comment above, improving efficiency (by die shrinks) simply permits one to increase the speed, usually up to just shy of reliability problems. I think by "efficient" you just mean cooler; your theoretical "100% efficient machine" is not at all; it doesn't do anything (we care about ops/joule (or ops/sec/watt)) \$\endgroup\$ – Nick T Jun 6 '13 at 20:00
  • \$\begingroup\$ @NickT: Faster operation requires the use of higher voltages than would be necessary at slower speeds. Assuming fixed parasitic capacitances, higher voltages will increase the amount of charge "lost" per switching event, and increase the amount of energy represented by each unit of charge. \$\endgroup\$ – supercat Jan 10 '14 at 22:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.