Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

The answer is 'simple' it comes in four parts;

1) Transistors dissipate the most power when they are switching (this is because they transit their linear region). So the more they switch the more power they dissipate, thus the power dissipation of any transistor circuit is proportional to its switching frequency.

2) Silicon stops working reliably above 150 degrees C, this is due to thermal effects in the silicon overwhelming its electrical characteristics. One way to think about it is that at 150 degrees C the silicon lattice structure is vibrating so hard because of heat the electrons can't move anymore.

3) Lithography and manufacturing techniques have increased the number of transistors per square millimeter of silicon exponentially over the years so the number of transistors has been observed to double every 18 months or so.

4) The ability to channel heat from silicon to the outside air is limited by the thermal conductivity of silicon and the ceramics used to encase it.

So those four parameters create a 'box' which is sometimes called the 'design space.' If you build a chip that is inside the box it works, if one of the parameters goes outside the box it fails.

In the great Mhz race of 1998 clock rates were pushed up which caused heat dissipation to go up, and transistor counts were going up too, so you got an n^2 effect in terms of heat dissipation. The race was powered by consumers who used the single number to compare machines (spec wars). It was unsustainable.

The end of that war occurred when AMD introduced multi-core (two cpus for the price of one!) architectures. And Intel had the largest design failure in its history when it scrapped the entire Pentium 4 microarchitecture after realizing it would never get to, much less past 4Ghz as they had promised.

AMD proved that for user visible throughput, multiple cores could give you a better net gain than a faster CPU. This side stepped the n-squared problem by having twice the number of transistors but running at the same clock rate, so the heat only went up linearly rather than exponentially.

It was a pretty humbling moment for Intel.

That begat the 'core wars' where Intel and AMD have worked to give us more and more 'cores'. The heat problem was still there, but it was managed with transistor design since the frequencies were staying flat.

Recently, some new transistor designs and system micro-architectures, have combined with the inevitable flattening of performance gains from multiple cores (see Amdahl's law), to give us 'turbo-boost' type solutions, where only one core runs at higher speed (side stepping Amdahl) at the expense of down clocking or even turning off other cores (sidestepping the frequency component of power increases).

Another technology on the horizon (which used to be only for the military guys) is SoD or Silicon-on-Diamond. Diamond is a wonderful conductor of heat and so if you make your processor on a diamond subtrate you can pull lots of heat out into an attached cooling system. Get ready for 7" x 7" heat management assemblies though that attach to these things. Or alternatively a new ATX type form factor that includes something that looks like a power supply but is a chiller that attaches via tubes to the processor socket.



clock rates were pushed up which caused heat dissipation to go up, and transistor counts were going up too, so you got an n^2 effect in terms of heat dissipation

Back when Dennard scaling was in effect, transistor power decreased as density increased, keeping chip power under control. The root of the problem is the end of Dennard scaling.

AMD proved that for user visible throughput, multiple cores could give you a better net gain than a faster CPU.

I don't know if that's true. AMD's success in 2003-2005 was mostly with single-core processors; I think all AMD showed was that a well-designed core and memory system beats a bad one.


It wasn't the multiple cores that allowed AMD to steal market share from Intel. It was the original Athlon chips around the turn of the century. They were the first chips that ran at a lower clock speed but produced far superior benchmark numbers in just about any benchmark. Combine that with the much higher prices that Intel charged for inferior products and AMD had a great run until Intel ditched the Pentium 4 line.

http://en.wikipedia.org/wiki/Athlon

Dual core chips came along much later.


The original Athlon was all about clock speed and the race past 1Ghz. The P4 quickly surpassed it, though, and left the Athlon competing on the basis of cost (P4s required RDRAM, which was much more expensive.) In the fall of 2001, the P4 hit 2Ghz and added DDR SDRAM support before the Athlon XP showed up at 1.5Ghz, and the Athlon XP was never really able to clearly beat Intel's fastest. It wasn't until the Athlon 64 that AMD was able to win across the board with significantly lower clock speeds.


So does point 1) mean that it takes too much energy to run a faster CPU, because power is always leaking away?

Also curious--are transistors small enough now that they can experience weird quantum effects?


No, point 1 means that 'active' power (which is the power the system dissipates when it is actively switching things) limits the maximum clock rate. Higher clock rate, more heat is generated, heat can't get out of the chip fast enough so chip temperature rises, transistors fail.

Leakage current, that is to say the current that flows through transistors because they cannot be fully turned 'off' is a sort of power 'floor.' If you use the 'water flowing' model of heat transfer (its not useful for simulation but it can explain behaviors) the leakage current results in a fixed amount of heat being dumped into the chip. Active power dissipation gets added to that and the combination is the total power created. Chip designers build systems to transfer that heat into an ambient cooling fluid (typically air) and away from the chip. You can do that by blowing air over a large surface area (as room temperature air goes by, it becomes warmer and the heat sink becomes cooler).

The transistors have always been small enough for 'weird quantum effects' the most well known is quantum tunneling of electrons into the gate area. Those effects are known and designed around but so far have not prevented the shrinkage of transistors (increasing the areal density of same).


Wow best post ever. Thanks man/woman!


Thanks for the thorough write-up!


I really want to start seing 128 or 256 bit processors... but with brand new architectures, putting aside the x86 and 64x86 architecture, with a huge instruction set like gpus... I wonder how much time could someone need to break sha1 in a 64 core 256 bit gpu alike processor.




Consider applying for YC's Fall 2026 batch! Applications are open till July 27.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: