hobold (hobold.delete@this.vectorizer.org) on September 30, 2013 3:00 am wrote:
> David Kanter (dkanter.delete@this.realworldtech.com) on September 29, 2013 11:39 am wrote:
> [...]
> > That being said, I think the move from 28nm to 20nm or 16nm will be very tricky
> > because the process technology won't be as much help as everyone thinks.
> >
>
> It seems to me that the end of Moore's Law is something that no one wanted to
> see coming, but everybody will say "I told you so" as soon as we're beyond.
>
> IMHO the benefits of Moore's Law will end quite a bit before transistors stop shrinking.
> In the good ole' days, a shrink of structure size gave us three nice consequences:
>
> 1. lower energy consumption
> 2. higher switching speeds
> 3. lower cost
>
> Back when Pentium 4 was the first major CPU to be made in 90nm, the first
> of those three consequences started to vanish. (It seems to me the lion's
> share of design efforts these days is focused on compensating for this.)
>
> Today, number two on the list is on its way out.
>
> Number three will be next. I suspect that the difficulties and
> cost of moving to 450mm wafer size is foreshadowing this.
>
> Nevertheless, transistors could continue to shrink a bit further, and nominally support the letters of Moore's
> Law a bit longer. I am just not sure if it will actually be done when there is no real benefit to be had.

So far, the benefits haven't completely stopped, but just slowed down.

Process shrinks in the 1990s gave huge benefits, sometimes doubling performance
by the time the process generation was complete and the design had been tweaked
to take advantage of the new process.

Intel processors make a good example. In mid 1997 Intel introduced the
"Klamath" Pentium II processor, fabricated on a 350nm process, at 233-266MHZ.
By late 1997 the fastest bin was up to 300MHz. Going to the 250nm process in January 1998 initially only sped up the design to 266-333MHz (with "Descutes"), but by late 1998
that design was up to 450MHZ, 50% faster than 350nm's fastest. Meanwhile, intel
introduced the Pentium III design with with a similar pipeline, Katmai, that
started at 450-500MHZ February of 1999, and eventually made it to 600MHz by August.
The next generation "Coppermine" processors, on 180nm, started at 500-733MHZ in October
1999 and eventually made 1.3GHz by 2001. (Note that the pipelines of these processors
are similar, so MHz is a reasonable proxy for process performance.)

Performance benefited from the reduced capacitance of smaller features.
Power per transistor not only benefited from the reduced capacitance, but it got a larger
benefit from the fact that the voltage was reduced with every process generation.
But, this stopped once voltage hit 1V. Process shrinks still helped power per transistor,
but more closely spaced lines meant higher capacitance, and hence power, per area and MHz.
The continued exponential increase in transistors meant dramatically increasing chip
power, which needed increasingly sophisticated design techniques to combat.

Power was one wall, but transistor speeds don't increase with each process shrink as
much as they used to, and wiring can sometimes actually get slower. So, while a leading
edge process still buys improved performance, the degree of improvement is much smaller
than it used to be.

Meanwhile, "Moore's second law", or "Rock's law", says that the cost of a fab doubles every 4 years. Evetually, the increasing cost of a new process generation will exceed the
decreasing benefit of process scaling, and the scaling will come to a stop. Moore's
law is a type of bubble. Like many bubbles, the existence of the bubble is easy to
spot, and it isn't hard to tell that the bubble must burst at some point. Predicting
exactly when it will burst is much harder.