The Not-So-Cool Future

markmcb writes "Researchers at Purdue University and several other universities are looking to start work on a major problem standing in the way of future chip design: heat. The team is proposing a new center to consolidate efforts in finding solutions for the problem that is expected to become a reality within the next 15 years as future chips are expected to produce around 10 times as much heat as today's chips. The new center would work to develop circuits that consume less electricity and couple them with micro cooling devices."

Nothing new

[koreaman]

What this boils down to is "researches are looking at ways to make cooler chips." Well, duh, haven't they always?

Photonic chips?

[Mysticalfruit]

I thought the future of processors was going ot be photonic processors. I'm not sure if these will be producing any heat or not.

1kW?!

[AaronLawrence]

("ten times as much heat as today's processors")
I don't think that 1kW processors will be practical. Nobody is going to want to pay to run that, and nobody will want a heater running in their room all the time either.

I'd say that they should be looking to limit it to not much more than current figures (100W) - maybe 200W if we are generous. After that it gets silly.

But can you make a cluster of them...?

[ites]

Not a joke.

The future is multi-core / multi-CPU boards where scaling comes from adding more pieces, not making them individually faster.

Yes, chips will always get faster and hopefully cooler, but it's no longer the key to performance.

Re:Photonic chips?

[Have Blue]

Everything that performs work produces heat. This is what we mean by "nothing can be 100% efficient".

Various solutions

[jd]

One "obvious" solution to the chip heating problem would be the following:

• Have a thin layer of some liquid like flourinert over the chip surface. It just has to conduct heat well, but not electricity.
  
• Put a Peltier device in contact with the top of the liquid. Peliters are metal, so that's why you want the electrically insulating layer.
  
• Have the top layer of the Peltier device double as a cold-plate.
  

This would let you get all the benefits of existing tried-and-tested cooling methods, but would eliminate the bugbears of the chip's casing being an insulator and the possibility of condensation screwing everything up.

A variant on this would be to have the chip stand upright, so that you could have a cooling system on both sides. The pins would need to be on the sides of the chip, then, not on the base.

A second option would be to look at where the heat is coming from. A lot of heat is going to be produced through resistance and the bulk of chips still use aluminum (which has a relatively high resistance) for the interconnects. Copper interconnects would run cooler, and (if anyone can figure out how to do it) silver would be best of all.

A third option is to look at the layout of the chips. I'm not sure exactly how memory chips are organized, but it would seem that the more interleaving you have, the lower the concentration of heat at any given point, so the cooler the chip will run. Similarly for processors, it would seem that the more spaced out a set of identical processing elements are, the better.

A fourth option is to double the width of the inputs to the chips (eg: you'd be looking at 128-bit procrssors) and to allow instructions to work on vectors or matrices. The idea here is that some of the problem is in the overheads of fetching and farming out the work. If you reduce the overheads, by transferring work in bulk, you should reduce the heat generated.

Re:Expect to see Asynchronous Processors instead

[dfghjk]

"And as a general rule, RISC processors are more efficient than CISC processors running at the same clock speed"

Where did that "general rule" come from? It's nonsense.