Intel Leads In Process. But By How Much?
Intel is ahead of the semiconductor industry in process technology but by how much?
Ed McKernan at SemiWiki writes: “They (Intel) believe they just need to get to 14nm production with finfets in 2014 and then they will be all alone with a 4-year process lead.”
And Malcolm Penn, CEO of Europe’s leading semiconductor analysts
company, Future Horizons, says: “TSMC are still lagging Intel by a
couple of years.”
However Intel’s process plans have not been totally smooth.
“Intel was supposed to have 22nm at the end of this year or Q1 2012. Now
this has been moved forward to sometime in 2012,” says statistical
variability expert Professor Asen Asenov of GlasgowUniversity and CEO
of Gold Standard Simulations (GSS).
At TSMC, the only other company able to compete with Intel on process,
Maria Marced, President of TSMC Europe, says: “The plan is to introduce
the first version of 20nm in the second half of 2012.”
And Lance Howarth at ARM says: “The expectation is that we’re a year
away from 20nm as a production technology.” ARM and TSMC have taped out
a 20nm chip.
So if Intel, TSMC, Asenov and ARM are correct, both Intel and TSMC will
be making ICs on 20/22nm processes next year.
That suggests an Intel lead measured in months rather than years, but there
are other considerations.
Mike Bryant, CTO of Future Horizons, points out: “Intel transferred the FinFET cell libraries to its Ivy Bridge design team in 2010 and so they were able to begin layout and simulations of the first device soon after. Although it is impressive that TSMC have a 20/22nm node device in design this is only a little further on from the SRAM test chips Intel used to first test the 22nm FinFET process in 2009.”
“The real indication of the lag between them will be when TSMC’s prime
Customers receive libraries allowing them to begin layout,” adds Bryant, “this is the
downside of the foundry model as Intel can risk sending a cell library to internal design teams in a state that TSMC could not send to its external customers.”
If Intel had cell libraries with its design team the year before last, and
TSMC has yet to deliver cell libraries to its lead customers, then Intel’s
lead on TSMC could well be two years.
But this is not the whole story. These days the number used to designate a
process is just that – a designator. It is not a measurement of gate length
with the shorter the length meaning the better the performance
characteristics of the process. These days the number ’22nm’ ’20nm’ etc is
just a designator, a description, and the value of a new node depends simply
on the advantages it delivers.
TSMC’s Marced says: “Compared to 28nm, the 20nm process is expected to
deliver a 25% improvement in power consumption, a 15-20% improvement in
performance and a 1.9x increase in density.
Can Intel say that their 22nm process will deliver that sort of advantage over their 32nm process?
Not according to ARM’s Howarth who reckons: “Their (Intel) vision is that finfet comes in at 20nm but the advantage of finfet will be marginal.”
Bryant reckons: “Regarding the relative performance of various technologies,
they all give roughly similar gains in power/speed/density and it is just a
choice of where on the power/speed curve you wish to lie.”
“Despite the best wishes of marketing departments, there are no magic
bullets that make one inherently superior over another,” adds Bryant, “but
the key to this node is the use of fully depleted channels. FD-SOI and
Intel’s TriGates have this and if TSMC have remained with planar then my bet
is they have achieved this in an as of yet unannounced manner, either by
using SOI wafers or by engineering the area under the channel.”
The best guess, as of now, is that the TSMC and Intel 22/20nm processes will
have similar characteristics.
Much has been made of Intel running 14nm ICs in its labs. The chips are not
due for production until 2014 – and it is normal to have three years between
generations i.e. 32nm in production – 22nm starting in 2012 – 14nm in
2014/15 with test chips running in labs about three years before
they move into fabs.
TSMC very likely has test chips on a similar process running in its own labs.
Volume production is, as always, the only certain-sure differentiator.