Intel’s former chief architect: Moore’s law will be dead within a decade

[humbug]

http://www.extremetech.com/computin...itect-moores-law-will-be-dead-within-a-decade

.............That Moore’s law will continue until 7nm or 5nm is actually extremely reasonable. I’ve heard other engineers speak of being dubious about 10nm and below. But the problem is simple enough: With Dennard scaling gone and the benefits of new nodes shrinking every generation, the impetus to actually pay the huge costs required to build at the next node are just too small to justify the cost. It might be possible to build sub-5nm chips, but the expense and degree of duplication at key areas to ensure proper circuit functionality are going to nuke any potential benefits.....

I think we can already see this happening, Bloomfield / Lynnfield to Sandy Bridge (45nm to 32nm) was a significant and tangible improvement in terms of thermals, power consumption and performance.

Sandy Bridge to Ivy Bridge (32nm to 22nm) was almost no performance increase but with tangible power envelope improvements, tho I would suggest that 3D Transistors play a significant part in that, not so much the DIE shrink.

Some optimisation saw Haswell Improve on Ivy Bridge, a bit.

Going forward from here node shrinking may only see marginal returns, pretty soon perhaps nothing. We are already seeing diminishing returns.

I think that is also very clearly demonstrated by the flat lining of that graph.

 

[Orangey]

Is there any chance AMD can get back in the race when we change over from silicon?

 

[humbug]

Is there any chance AMD can get back in the race when we change over from silicon?

AMD need to make architectural changes, the 32nm DIE on Piledriver is not the problem, going to 28nm is probably not going to make a great deal of difference 'on its own' so Steamroller needs to be a much improved architecture, which according to apparently leaked Diagrams and DIE photographs it is, its doubled literally everything feeding the cores.

It should be said tho, those are rumours. AMD are pretty tight lipped on SR.

 

[shadow_boxer]

The limits of silicon are being approached. I totally agree that AMD need a new architecture with SR but even so the gains with new generations look like they're going to be pretty small until we see silicon itself being replaced perhaps.

 

[Suarez7]

I said this just the other day that we're reaching the limits on what silicon can do, unfortunately we'll be stuck with it for years to come. The prices of 10nm and below chips will be ridiculous.

 

[panyan]

unfortunately we'll be stuck with it for years to come.

why do you think that?

I was reading only the other day that Graphene is being looked into with a theoretical max frequency of 427GHz or something ridiculous like that

 

[Gripen90]

So I guess the problem will also reach GPUs albeit a bit later ?

 

[Suarez7]

why do you think that?

I was reading only the other day that Graphene is being looked into with a theoretical max frequency of 427GHz or something ridiculous like that

There was a program about it a few months back and they said it would be unlikely graphene would be used for computing chips for years to come unless they can find a way to completely switch the transistor off.

http://www.bit-tech.net/news/hardware/2011/01/21/ibm-graphene-wont-replace-silicon-cpus/

So I guess the problem will also reach GPUs albeit a bit later ?

Correct, it will affect everything that uses silicon. Apart from B00B implants.

 

[humbug]

So I guess the problem will also reach GPUs albeit a bit later ?

Because of the way fabrication works 28nm to 20nm is actually (as near as makes no difference) halving the size of the DIE with the same number of Transistors.

So theoretically a GTX 770 (1536 SP's) to a GTX 870 could have 3072 SP's, or a 7970 (2048 SP's) to 9970; 4096 SP's and maintain the same DIE size.

BUT, using the same architecture shrunk to 20nm it may still have anything upto twice the power consumption if there are no energy efficiency returns from the DIE shrink.

20nm wafers are also 40% more expensive.....







So are we going to see a doubling of GPU from 28nm to 20nm?

IMO, Hell no! from AMD we might see a 3000 SP GPU and a 25% smaller DIE, and the same again from Nvidia, GTX 870 would be = GTX 780

 

[funky_llama]

There was a program about it a few months back and they said it would be unlikely graphene would be used for computing chips for years to come unless they can find a way to completely switch the transistor off.

http://www.bit-tech.net/news/hardware/2011/01/21/ibm-graphene-wont-replace-silicon-cpus/



Correct, it will affect everything that uses silicon. Apart from B00B implants.

This was slashdotted a few days ago. Maybe not time to discount graphene yet. http://www.technologyreview.com/view/518426/how-to-save-the-troubled-graphene-transistor/

400Ghz yes please

 

[Suarez7]

This was slashdotted a few days ago. Maybe not time to discount graphene yet. http://www.technologyreview.com/view/518426/how-to-save-the-troubled-graphene-transistor/

400Ghz yes please

All that says is that they think they've found a solution, till they actually prove it I'll still say graphene chips for the mainstream are years away.

 

[humbug]

I'm with Suarez7 on Graphene chips, the amount of engineering, testing and red tape these things need to go through before the concept is even fully prove let alone approved means its an absolute certainty that we will not be buying any for many years to come yet.

 

[funky_llama]

We wont be seeing them for another decade at least thats for sure. Will certainly make things interesting though.

 

[humbug]

We wont be seeing them for another decade at least thats for sure. Will certainly make things interesting though.

Yeah, provided they are combined with another concept (magnetic switching transistors) what will happen is these things become so small you will have to power of a desktop in something the size of a pinhead and use almost 0 power while producing no heat.

 

[Hakanese]

And suddenly quantum computing becomes a little bit closer

 

[tbh]

So I guess the problem will also reach GPUs albeit a bit later ?

Nvidia have already spoken of their concerns in this area:

http://www.extremetech.com/computin...y-with-tsmc-claims-22nm-essentially-worthless

There was a program about it a few months back and they said it would be unlikely graphene would be used for computing chips for years to come unless they can find a way to completely switch the transistor off.

That's why graphene is so two thousand and late and silicene is The Next Big Thing.

 

[JonSmith]

Still... Think of the heat caused by switching a component at 400GHz.

 

[tbh]

The perf/clock line on the graph in the original post doesn't seem quite right, it flat lines after the Pentium 4 where in reality it should continue to climb until the Core 2 Duo and then gradually but not completely level off through Nehalem, Sandy Bridge etc.

 

[HangTime]

^Couple of things to bear in mind:

1) If you look at the dots on the chart rather than the line of best fit, Core 2 is significantly above P4 considering it is a logarithmic scale (typically such graphs will be much flatter than you expect compared to linear scales). Things like Nehalem and Sandy Bridge aren't even plotted. The fact there's only one data point plotted after late 2003(?) compared to dozens prior to it kinda skews it a bit.
2) Performance per clock is shown as being largely flat from 2000 through to 2004 which to be fair on the Intel side is fairly accurate - the progression from Coppermine > Williamette > Northwood > Prescott was largely driven by clockspeed increases, heck P4 Williamette was arguably slower than P3 Coppermine clock for clock and again that is illustrated by the slight dip in the dots on the graph.

 

[HankScorpio]

I think Moores law will be maintained as long as humans continue to invent and innovate. If you look at ancient Greek hand crank machines/early computers and compare the graph from then until now, it's pretty short sighted to say the law is coming to an end because the silicon method is showing its limitations. Graphene/biological and quantum computing maybe decades, even centuries away but the trend should still go upwards. Even if it flatlines for the next for years.