AI: Future of Humanity
Sphere of Interest
Stanford Brain Lecture Notes
The RX Project:
CV Biblio (1985)
Index of Essays
Robotics, and AI
Health & Biotech
Earth Wisdom: Universe
Be Saved by Bob!!!
(And Other Balms )
Are Fats Killers
The Mystery of CONSCIOUSNESS
Who, What, When?
Consciousness & Brain
Near Death Experiences: In the Desert With Pim Van Lommel
Is the UNIVERSE
Fine-Tuned for Life?
Neuron Videos Say
Forget Realistic AI
EUV 2014 - Future of Moore's Law
BAM: Brain Activity Map of Spikes
What is Watson?
AI Overlord or Tool?
SETI: Search for Extraterrestrial Intelligence
KEPLER Seeks Earth-like Worlds
STEVE PINKER in the Amazon: photos
Billion Year Plan:
CONSCIOUSNESS as Global Resonance
CT Scan: Yes!
Book Review: TRANSCEND
Create a Mind
Does Drug X
Works in Monkeys!
The Future of Moore's Law
At the heart of all the research I follow is Moore's Law (the doubling
of transistor density on a chip about every two years).
While raw computer power alone will not create transhuman AI,
large scale brain simulations, or genomics miracles,
it provides a fundamental, enabling substrate for those enterprises.
I write this article using a 2012 Intel Ivy Bridge cpu designed
at the 22 nm design node. (The design node refers roughly to
the size of the smallest features on the chip - usually the gate length.
Pitch refers to the distance between successive transistors (like the
distance between fence posts) and may be 4 times the node size.)
For comparison, a bacterium is about 1000 nm and a red blood cell is
7500 nm. The width of a human hair is about 70,000 nm.
One nm = 10 angstroms = 4 * the width of a single silicon atom.
Each (single) strand of DNA is about 2 nm in width (but centimeters long).
The next cpu die shrink from Intel will be their Broadwell cpu
scheduled for release in 2014 at the 14 nm node, and in 2016
Intel plans to release its Skymont cpu at 10 nm. However, the
14 nm node may actually be more like 16 or 17 nm. (EUV is too late
for 14 nm and may even be too late for 10 nm, but see this
Broadwell announcement from Intel's Paul Otellini in 2013.)
When will Moore's Law hit a brick wall? For CMOS
transistors, it will be within a decade -but can it even go that long?
I read every new edition of the ITRS Roadmap on Emerging Technologies.
CMOS transistors will be replaced by one of several other promising
methods for storing bits and logic, eg graphene/ carbon nanotubes and/or
bottom-up directed self-assembly.
Meanwhile, what are the prospects that the semiconductor industry
will be able to maintain the every two year tick-tock of Moore's Law?
For the past several years all wafer fabrication has used 193 nm light
even for my 22 nm Ivy Bridge cpu chip. It's astounding - it's like using
a paint brush for a house to do a fine portrait. It's possible, but the many tricks
for doing that (water immersion, multiple patterning, fine stepping alignment)
are almost played out, although they may make it possible to get to 10 nm.
A bright ray of hope for allowing Moore's Law to continue to < 10nm
is extreme ultraviolet light (EUV). The industry has standardized on
13.5 nm EUV, and it's so difficult to generate and focus that only one
key player remains: Cymer, based in San Diego. In 2012 Cymer was
purchased by Netherlands-based ASML, a world leader in
photolithography. And, more importantly, to share the risk
of this difficult R and D, ASML negotiated a 15% buyout of its own stock
by Intel. Taiwan Semi and Samsung also bought large positions.
This Cymer video explains how their LPP (laser-produced plasma)
light source works. It takes heroic effort to get enough power as
this video on pre-pulse shows. The bottom line is that EUV litho is
working; they have shipped six of their pre-production NXE:3100
machines, and customers have exposed more than 30,000 wafers.
Furthermore, their production EUV machine, the NXE:3300B, is being
assembled and 11 orders have been received. EUV volume production
will begin in earnest in 2014. Current status of EUV as of 17 Apr 2013
was shown in ASML's Q1 2013 press release and here, note espec. slides 22/23(!).
Cymer NXE:3300b Vessel
But note carefully that in 2014 only "critical layers" of wafers
may be printed using EUV. This is because the power of the EUV source
is still too weak to allow high wafer throughput. Here Cymer describes
their process for laser generation of 13.5 nm EUV light. Although
the process begins with a 30,000 watt CO2 laser used to explosively
create EUV from 30 micron tin droplets, the power of the EUV at the
intermediate focus is now only about 55 watts. And, by the time
the EUV has been focused by a series of reflecting mirrors only about
a hundreth remains to expose the photoresist-coated wafer.
(But, CYMER/ASML hopes to hit 250 watts and 126 wafers per hour by 2016.)
So, the current machines work, but more intense EUV is needed
The extreme challenges posed by EUV and by next generation lithography
before the industry can economically produce all layers of 14 nm and 10 nm
wafers with EUV. Whether and when that will occur has been the subject of
intense speculation. See these articles by Vivek Bakshi and this YouTube by
Chris Mack. Meanwhile, Intel with its multibillion dollar investment in ASML
is trying to minimize the risk of failure. Intel's Mark Bohr has stated that
Intel has a path down to 10 nm even without EUV (by using quadruple
patterning with 193i), but Intel hopes that won't be necessary. Quadruple
patterning is at best four times as slow as single exposure. Here is an Oct 2013
update from ASML CEO Peter Wennick that addresses this trade-off.
were just summarized by IBM's Gary Patton and others at the 2013 Common Platform
Technology Forum (talks are online).
I live across the street from a 3 km long linear electron accelerator
(SLAC) that has been redesigned as a source of coherent light (LCLS) for
imaging of single nm biologic molecules and advanced materials.
Another accelerator at Lawrence Berkeley National Lab (LBNL) is being used
to create synchrotron radiation, which is being used in a collaboration with
the semiconductor industry. The Sematech/LBNL collaboration has created tools
for testing EUV for wafer fabrication.
Despite the availability of low cost
synchrotron radiation (Lyncean), it appears that this may only be used
for metrology, eg defect analysis, rather than for fabrication. Nonetheless,
my SLAC physicist neighbor and fellow juggler, Gennady Stupakov, suggests
that tunable free electron lasers may solve the problem of increasing EUV energy.
EUV progress (and challenges) on semimd.com . Here is a September 2013 update
on EUV - Taiwan Semi will attempt to use it for 10 nm production in 2016.)
Apart from EUV another former bright hope for continuing Moore's Law was
e-beam lithography. This is still being pursued by a few companies but
to be cost-competitive requires the parallel use of millions of simultaneous
electron beam writers. Without further progress, this is apt to be a niche technology.
(Electron microscopy (SEM / TEM) is a mainstay of biology research, but electron beam
writers are apt to be a niche fab technology.)
Below 10 nm a number of methods for directed self-assembly (DSA)
are being explored. (This video shows DSA for MEMS self-assembly. However,
DSA usually refers to the use of block copolymers to create fine
patterning between coarser lines printed with lithography as in this video, and in
this article. DSA is now creating a huge buzz among litho players.
But, that's a story for another day.