Semiconductor industry set to abandon Moore's Law.

[CAPSLOCKSTUCK]

The semiconductor industry, which has used the observation made by Intel co-founder Gordon Moore as its target development in recent decades, is now set to abandon the approach.

The move is thought to be a reflection that companies are struggling to keep up with the pace of innovation required to cram ever more transistors onto a finite space. Among the most limiting issues has been the heat generated as more and more circuitry is jammed onto a silicon chip

According to the International Technology Roadmap for Semiconductors, the industry body that sets out the direction it will take, is holding its final meeting in Atlanta, Georgia, this week as it prepares to lay out its plans for between 2015 and 2030.
It has said it must 'define new drivers to help it stay on a path of productivity and profitability'.
A meeting of the organisation next month will 'for the first time drive a top-down system-driven roadmap framework'.
This will follow an approach that is being called More than Moore – which will aim to develop chips appropriate for applications like smartphones and supercomputers rather than simply trying to aggressively improve performance and meet cost requirements.
A paper published on the ITRS website said: 'Historically, the ITRS has used metrics such as transistor density, number of cores, power, etc, to roadmap technology evolution of integrated circuits.
'These metrics are essentially driven by the physical-dimension scaling as predicted by Moore's Law.
'However, new requirements from applications such as mobility, datacenters, etc require a new, system-level roadmapping approach, as these applications imply roadmaps for system-level metrics.'

Moore's Law predicts that the number of transistors on a chip will double ever two years. This has proven accurate for the past few decades (illustrated above) but it is now slowing down

 

[rtwjunkie]

Thanks for the post! I wouldn't be surprised if it gets moved to Science and Technology, though.

 

[trog100]

they have decided that chips are powerful enough and tomorrows priority will be to make them more energy efficient and generate less heat as opposed to go faster.. he he

not really scientific more philosophical in a way..

looking at all the huge cooling machinery inside my mid tower case i think i might right agree..

something the size of a house brick being needed just to stop a tiny fingernail sized chip from frying dosnt make a lot of sense..

trog

 

[Vayra86]

Makes sense that this would someday end, process node shrink is getting harder and harder so this is the only logical consequence.

We have already moved into the realm of efficiency / coding improvements to net our biggest gains in performance. DX12, Mantle, also point towards that.

 

[MIRTAZAPINE]

Thanks for scientific papers. Looks like another long read for me in the weekends. The problem with moore law is that it is an exponent and it have limits for it growth in our finite world. The problem of moore law is like the story of wheat and chessboard problem. The story is someone in ancient time when asked for a reward from the king said that to give one grain of rice in one day. Then the next day double the grain of rice. The previous day rice would be added with the new day. Then the next day double again. During the start the effect is not apparent yet until you grow to a huge number where doubling is no longer feasible. After the 32th square is past.

I am sure Gordon Moore do see the exponential growth problem then. It is only a matter of time.

 

[Bill_Bright]

The semiconductor industry? Says who?

First off, the ITRS, contrary to how important they think they are, do NOT "set the direction" the semiconductor industry takes! It merely maps the evolution the industry has taken. Not the direction it is taking or will take. Big difference!

Note the ITRS's own disclaimer,

"The ITRS is devised and intended for technology assessment only and is without regard to any commercial considerations pertaining to individual products or equipment".

As far as Moore's law being dead or alive really depends on who you talk to. I note the law has evolved just as the CPU has. For example, originally, it was based on a 18 month time frame and that has been extended to 24 months.

Note the January 11, 2016 article by Mark Stettler and Shesha Krishnapura from Intel where they write, Moore’s Law – Not Dead – and Intel’s Use of HPC to Keep it Alive.

And note the problem is not really heat "generation". As densities increase and the gates get smaller and smaller, the voltage necessary to "flip" and "flop" the gates decreases too and that allows for today's CPUs to consume much less power, be much more efficient and thus generate much less heat in the process for the amount of work being done.

The problem with keeping the law alive is based just the physical limitations of the silicon atom - that is the physical width of the gate in the IC. As more gates are jammed into the same amount of space, the gates get smaller and smaller. A silicon atom has a specific physical size and these gates are approaching single molecule dimensions.

There are several other factors that come into play when pushing electrons (voltage) through any circuit. There is (yet) a perfect conductor. That is, there is always resistance in every circuit, no matter how small or efficient a circuit may be. To overcome that resistance, you must increase the voltage.

There are two factors that will cause voltage to "arc" - jump across a gap (the gate in this case). One is the distance across the gap, and the other is the "potential" of the voltage. If you increase the voltage, there is greater chance there will be an arc. If you decrease the gap, there is a greater chance there will be an arc. So there is a huge challenge to provide enough voltage to operate the gate but not enough to jump across an atom wide gap. Keeping Moore's Law alive will be a challenge and going vertical (3D) will help, for now. But the real solution is coming up with something other than silicon. Then Moore's Law will simply be retired, not dead (or it will evolve to address the materials used in place of silicon).

 

[the54thvoid]

There are two factors that will cause voltage to "arc" - jump across a gap (the gate in this case). One is the distance across the gap, and the other is the "potential" of the voltage. If you increase the voltage, there is greater chance there will be an arc. If you decrease the gap, there is a greater chance there will be an arc. So there is a huge challenge to provide enough voltage to operate the gate but not enough to jump across an atom wide gap. Keeping Moore's Law alive will be a challenge and going vertical (3D) will help, for now. But the real solution is coming up with something other than silicon. Then Moore's Law will simply be retired, not dead (or it will evolve to address the materials used in place of silicon).

I assume this is the the issue with overclocking (over volting in current gpu chips). I read from EVGA's Vince Lucido that the Kingpin (Maxwell arch in general) doesn't take to voltage very well. More voltage increases heat which causes 'ripple' in layman's terms. This instability of voltage in a circuit causes problems. So as things get smaller it is harder to control, as you say, the containment of the applied current, without keeping things very cold. This explains why maxwell and other chips clock so much higher under super cold conditions - current stability (as cold reduces atomic movement)?

 

[Bill_Bright]

I assume this is the the issue with overclocking (over volting in current gpu chips).

Not really the same thing because the protection circuits [hopefully] cause a shutdown before then. Arcing is a spark and inside a processor from one conductor to another (or from a "high" to a "low" or to ground). It might as well be a hot bolt of lightning - microscopically speaking. That is, arcing will result in permanent damage - much like zapping a CPU with ESD does.

No, not even in laymen's term is that "ripple". Ripple is like the little waves created when you throw a rock in the water of a still pond. The flat surface is the "DC" voltage and the "ripples" are little "AC" fluctuations riding that DC.

In laymen's terms, that increase in heat causes "instability" - as you noted.

(as cold reduces atomic movement)?

And cold is also the absence of heat - which, as noted by the last line in my sig, is the bane of all electronics.

 

[Frick]

So what's the punishment for not obeying?

 

[trog100]

i think the latest intel skylake shrink kind of makes the point an end of line has been reached.. the law of diminishing returns always kicks in eventually..

when the first shrinks occurred the smaller chips always needed less volts to run.. now it dosnt seem to be the case.. i see 6th generation skylake chips needing roughly the same voltage and generating roughly the same heat as my 4th generation 4790K chip does.. the max wattage goes up from 88 to 91 and the clock speed goes down.. not exactly what i expect from a shrink..

trog

 

[Bill_Bright]

So what's the punishment for not obeying?

You just get a processor that is more efficient and generates less heat for the amount of more work it is capable of doing - but sadly, does not have twice the number of transistors on the die.

i think the latest intel skylake shrink kind of makes the point an end of line has been reached.. the law of diminishing returns always kicks in eventually..

I think it shows it is getting there.

If you look at the i7-6700K compared to the i7-4790K there are certainly a lot of similarities but there are several differences too. The most significant may be the transfer/second speed. The Skylake supports 8 GT/s on a faster bus while the 4790K "only" does 5 GT/s on the slower bus.

I note too the Skylake i7 has 1.75 billion transistors stuffed in 122 mm² while the Haswell has "only" 1.4 billion transistors in a die 177 mm². While not "twice" the number of transistors, 350,000,000 more transistors in 2/3 the space is clearly a significant shrink there.

 

[yogurt_21]

I note too the Skylake i7 has 1.75 billion transistors stuffed in 122 mm² while the Haswell has "only" 1.4 billion transistors in a die 177 mm². While not "twice" the number of transistors, 350,000,000 more transistors in 2/3 the space is clearly a significant shrink there.

1.25 the amount of transisters spread out over .66 of the area = 1.89 times the transisters per area... hence the meme.

 

[FordGT90Concept]

Devil's Canyon and then again again with Kaby Lake are proof enough this is true. I think there will be a slump until the industry decides on where to go from here (silicon-germanium probably) and then Moore's Law will resume until that substrate is exhausted too.

 

[Bill_Bright]

1.25 the amount of transisters spread out over .66 of the area = 1.89 times the transisters per area... hence the meme.

Not sure I understand your point. Are you agreeing or disagreeing? That is, do you not agree that 1.89 times is pretty close to 2 times? And that that is a pretty significant shrink?

 

[yogurt_21]

Not sure I understand your point. Are you agreeing or disagreeing? That is, do you not agree that 1.89 times is pretty close to 2 times? And that that is a pretty significant shrink?

what part of "close enough" wasn't clear?

 

[Bill_Bright]

I don't see any reason to be snide about it. Your "image" for close enough could have easily be taken as sarcasm. Sorry if asking for clarification put you out.