Moore's Law Coming To an End

[Jccarlton]

The problem is not physics, it's cost. We can make smaller lines on silicon, but doing that is becoming more and more costly. I interviewed for a job at ASML and was able to see the next generation stepper machines. Perkin Elmer's first machine was the size of large desk and cost about 300k. Now these machines are the size of a bus and cost 500 million a pop. The engineering director and I both agreed that the end was probably about ten years or so away. That timeframe fits with this analysis:
http://nextbigfuture.com/2013/08/moores ... .html#more
I'm not sure what the consequences will be, but all good things come to an end.

[TDPerk]

Nowhere does that article provide useful metrics.

The cost of the machine is not relevant in and of itself. The issue would be how many transistors per dollar/day we're getting out of the newer machines.

Colwell literally provides nothing to justify his inference.

"But how about 20 per cent? How about 10 per cent? How far down are you willing to go and still think that you've got something you can sell?"

I think he doesn't understand economics in the slightest. If R&D for that style of machine is at a dead end, then the cost of the new machines starts to drop, the recently purchased machines start to amortize their costs, and with no other advantage to distinguish them, the 10% to 20% improved product will be sold, because that's how the economics of commodities work.

I find no justification to think the Memristor memory products and other than silicon substrates aren't going to keep it going to 2040. The production of physics limited processors is where it will stop, and not before the physics implementation becomes as yet unanticipated and clever.

[DeltaV]

I'm not sure what the consequences will be...

Skynet will be delayed until the orbiting quantum spintronic memristor neural chip autofabs are operational.

This gives us a little more time to develop countermeasures. Very little more time.

[hanelyp]

I doubt this will be an end to Moore's law. More likely just a hickup until the next major electronic technology takes hold.

[GIThruster]

The focus is shifting toward combined functions like embedded capacitance on ships. I doubt Moore has said all he will yet, but all good things do come to an end.

[necoras]

Most chips are still 2D. Build up. It worked for biology. Then you have to deal with heat buildup. Oxygen availability is only one reason for the amount of blood flow to the brain. There are plenty of ways to suck heat out of a chip. Liquid nitrogen works, but it's unwieldy. Graphene based chips may work better. There may be other non-silicon materials that dissipate heat better.

And none of the above even bothers with optical or quantum computing. Moore's Law may slow a bit, but it's not dead yet.

[Schneibster]

Actually, graphene. They've actually made a small, simple CPU containing operating registers, microcode, and an ALU out of it. The components are on the molecular scale. Saw it in Science Daily last week, IIRC. The process is tractable to existing semiconductor practice and to economies of scale.

Edit: Found the article, http://www.sciencedaily.com/releases/20 ... 132314.htm

I was wrong, the material was carbon nanotubes not graphene sheets.

[Schneibster]

And here's another, completely separate development: photonics. Not only that, photonics on-chip on microprocessors, using standard SOI CMOS and bulk CMOS techniques, full function transducers capable of both receiving and transmitting photonic signals, and of complementarily accepting and generating electrical impulses. And this report is on the occasion, not of their invention, but of an increase in their efficiency that will reduce the power consumption on multicore chips like the i7 potentially by as much as half. And they haven't even started trying to optimize the cores yet with photonics; just the communication between cores.

Furthermore they have a method for doing it on bulk CMOS which means the memory. Which means we're a hop skip and a jump from optical memory, which can operate at THz (terahertz, that is trillions of accesses per second).

Looks like Moore's law might have a hiccup and then start in again.

[MSimon]

Better software will help a lot:

http://www.ecnmag.com/blogs/2013/10/can ... moores-law

[MSimon]

Most chips are still 2D. Build up. It worked for biology. Then you have to deal with heat buildup. Oxygen availability is only one reason for the amount of blood flow to the brain. There are plenty of ways to suck heat out of a chip. Liquid nitrogen works, but it's unwieldy. Graphene based chips may work better. There may be other non-silicon materials that dissipate heat better.

And none of the above even bothers with optical or quantum computing. Moore's Law may slow a bit, but it's not dead yet.

Uh. The problem is heat production and transfer. 3D will make that worse.

[Schneibster]

Most chips are still 2D. Build up. It worked for biology. Then you have to deal with heat buildup. Oxygen availability is only one reason for the amount of blood flow to the brain. There are plenty of ways to suck heat out of a chip. Liquid nitrogen works, but it's unwieldy. Graphene based chips may work better. There may be other non-silicon materials that dissipate heat better.

And none of the above even bothers with optical or quantum computing. Moore's Law may slow a bit, but it's not dead yet.

Uh. The problem is heat production and transfer. 3D will make that worse.

Actually, with photonics, it won't.

Also there's internal cooling which hasn't even been properly explored yet. It's child's play to instantiate the channels at the foundry. You can fill them with anything from glycerine antifreeze to liquid nitrogen. You can connect them to pumps. None of this has been done other than in a few high-end applications.

And that assumes after redesigning with photonics there's enough heat left to dissipate to make it worthwhile.

[MSimon]

Most chips are still 2D. Build up. It worked for biology. Then you have to deal with heat buildup. Oxygen availability is only one reason for the amount of blood flow to the brain. There are plenty of ways to suck heat out of a chip. Liquid nitrogen works, but it's unwieldy. Graphene based chips may work better. There may be other non-silicon materials that dissipate heat better.

And none of the above even bothers with optical or quantum computing. Moore's Law may slow a bit, but it's not dead yet.

Uh. The problem is heat production and transfer. 3D will make that worse.

Actually, with photonics, it won't.

Also there's internal cooling which hasn't even been properly explored yet. It's child's play to instantiate the channels at the foundry. You can fill them with anything from glycerine antifreeze to liquid nitrogen. You can connect them to pumps. None of this has been done other than in a few high-end applications.

And that assumes after redesigning with photonics there's enough heat left to dissipate to make it worthwhile.

Coolant? Pumps? Do you know what that does to costs? And reliability? There is a reason it is only done in high end applications. Will the coolant stay in the system for 20 years? Will it need to be topped up? The real problem is chip design. Chips need to automatically idle when they are doing nothing. The GA144 is a good example of that. It is unclocked logic.

You may be unaware of this but making photons is a heat generating process. Absorbing them generates more heat. Sure you get electrons. But at the same energy as the incoming photon? And the electrons eventually lose their energy.

Sure - progress is being made. But I believe Moore's Law will plateau before these wonderful advances kick in.

Software may gain us a decade, unclocked logic another half decade.

None of the wonderfulness you mention is ready to be scaled up. So far they are small potatoes and cost a lot.

[paperburn1]

Question , anybody know what the life cycle of a heat pipe would be,? And could you make them small enough to be integral to the chip itself?

[ladajo]

There is a company in New Hampshire that came up with a thermally convective heat sink for high power computing in compact form factors. It was a heat sink with a radiator filled with a gel like sustance as I recall. This substance was fluid enough to thermally motivate itself around the loop.
The radiator was very thin and could fit behind the screen.

I'll have to see if I can find it again. I remember seeing a physical demo a few years back.

Found it:

http://www.aavid.com/

I think it was specifically related to this:

http://www.aavid.com/solutions/ex-surface

[hanelyp]

Part of the reason the CPU in your PC runs hot is the architecture. They take compact sequential instruction streams, addressing a smallish set of registers, that are portable across a wide range of devices. And the hardware compiles those into a pool of instructions that can be executed out of order and in parallel using a large set of registers on hardware that may vary extensively. Part of the speedup is "speculative execution", performing calculations before it is known whether the result is needed. The result is a much faster computer without the complications being exposed to the software developer, but a great deal of overhead to manage those internal instruction pools and registers.