Download - Optical computers
University of Ljubljana Faculty for Mathematics and Physics
Department of Physics
Seminar
Optical Computers
Author: Žiga Lokar Mentor: prof. Igor Poberaj
Ljubljana, 31.3.10
Summary:
This seminar will first explain benefits and problems with optical computers, then explain all vital
parts of a computer and how can they be replaced with optical equivalents. First, the seminar will
discuss feasibility to completely replace copper wires with optical fibres in motherboards, then
explain how is computation with photons possible. In the end, all optical storage and memory will be
discussed.
1. Contents 2. Introduction ..................................................................................................................................... 3
SOA ...................................................................................................................................................... 3
3. Backplane (motherboard) ............................................................................................................... 4
4. Processing unit ................................................................................................................................ 6
5. Storage ............................................................................................................................................ 9
6. Memory ......................................................................................................................................... 11
7. Conclusion ..................................................................................................................................... 13
8. References ..................................................................................................................................... 14
2. Introduction For computing, light is currently used for signal transmission, mostly for long distance
communication. It offers several benefits compared to electric signal transmission, most importantly,
much higher bandwidth. Furthermore, energy requirements are lower and photons also interact
weakly with electromagnetic field, meaning signal transmission is less prone to errors. The limiting
factor for long distance communication are currently electronic parts of the system. Signals need to
be cleaned of noise after some distance. This is usually realised with OEO (optic-electronic-optic)
devices; signal is converted to electronic one, cleaned and retransmitted as an optical signal. All
optical signal regeneration is also possible, experimentally demonstrated by various research groups
[1, 2]. Another desire that would also greatly speed up long distance communication is fast all optical
signal routing, also demonstrated [3].
SOA Semiconductor optical amplifiers are currently used mostly for signal amplification before detection,
where only amplitude is increased. Using nonlinear effects exhibited by these devices, signal can be
also reshaped and retimed. This is called 3R regeneration.
Image 1: Signal regeneration. 1R means amplification, 2R adds re-shaping while 3R adds re-timing as well. On x axis there is time, with amplitude on y axis.
Such regenerators work at much higher bitrates than electronic signal regeneration and there is no
need to convert signals. But their problem is small amplification and regeneration for amount of
energy needed. Only amplifiers are used in newer fibers, so there is no need to re-transmit the signal
every couple of kilometers, as it was needed in older fibers.
All demonstrated technologies show promise for fast all optical systems for long distance
communication, although there are still many serious problems how to scale the device and decrease
power consumption.
For computers, on the other hand, small system size is also required. Seminar will explain current
status of all-optical computers.
3. Backplane (motherboard) Backplanes are vital parts of computers, however, they are not known under this name in personal
computing. For personal computers, as well as quite many servers nowadays, term motherboard is
used. These terms are not identical; motherboards offer some integrated processing, as well as a
socket for the central processing unit, CPU. Backplanes are only used to route signals between
various add-on cards, and a CPU is just one of the inserted cards if there is a central processing unit
on the backplane at all. However, these differences are mostly semantic as both parts have identical
important role – interconnection between different components.
I will use term backplane during this seminar, as work in this area is based on providing fast
communication.
Optical backplanes consist of similar parts than long distance communication. Signal source is a
VCSEL, vertical-cavity surface-emitting laser. Benefit of these lasers is a low power requirement and a
very high efficiency, but their maximum power is very low as well. Signal is modulated using Kerr
effect, which will be discussed under processing unit. The signal propagates through optical fiber,
similar to the one used in long distance communication. No regeneration is needed, as distance is
short. On the receiving point, a photodiode is used to convert optical signal back to electric one.
Current demonstrations of optical backplanes do not plan to have optical signal routing, optics serve
only as a point-to-point link.
Major benefit of the optical backplane is its very high transfer speed. Current demonstrations exceed
10 Gbit/s/channel/fiber. Their major difficulty is its high cost of equipment, while there are some
other, physical limitations.
One of the problems is various sources of noise – noise on the laser, the detector and a fundamental
physics limit, shot noise. This effect is known to anyone investing some time in photography. When
the number of photons collected is small, this causes a large variation of the signal. Under
assumption there will be a Poisson distribution of photons and the detector has a constant
percentage to collect a photon, signal to noise ratio equals √ . This noise presents lowest possible
power consumption, as system is characterized by minimum signal to noise ratio. Electric systems
currently operate with less noise for short distance. Therefore, the advantage of decreased
interaction with environment is reduced due to power requirement of the connection. This only
holds true if the wire is very short, otherwise resistivity of wires greatly decreases performance –
signal attenuation due to distance travelled is much higher for copper wires than it is for optical
fibers.
Image 2: Shot noise on a photograph. http://en.wikipedia.org/wiki/Shot_noise; 21.3.10
Furthermore, frequency of light used in long distance communication fibers is 1550 nm. This
therefore limits minimum diameter of the wire to about 750, as light cannot travel through a
waveguide that is less than
thick. Further consequence is the imposed limit on the laser
dimensions, the very same 750 nm. This limit is much more important than the noise limit; it
presents a great difficulty to scale the system to on-chip communication, especially if routing is
needed.
Despite all difficulties, end result is in favor of photonics with higher transfer speed and lower power
need per GB/s, while cost problem remains. In HPC (high performance computing) connections
between 2 backplanes have been optical for some time now, due to mentioned benefits. Migration
towards optical backplanes has not started yet, although technology is ready and offers higher speed
than electronics. Shorter transfer length (even on chip optical communication) has been
demonstrated in laboratory, but practical use is not as close as it is for backplane optical
communication. [4, 5]
Image 3: IBM's idea of merging photonics, memory and processing on the same chip in different layers. Estimated by IBM to appear around 2018. [4]
For high performance computing, communication is therefore migrating towards photonics. Main
reasons are or need to become cost per bit for backplanes and communication between racks. For
on-chip communication, power per bit and maximum bandwidth is said to bring advantage to
photonic communication over copper wires.
For personal use, optical communication is further away. The first announced optical connection
between devices is Intel’s Light Peak, optical connection between a PC and other devices, such as
monitors, TVs, hard disk drives and so on. All-optical motherboards or smaller connections are not
expected yet.
4. Processing unit Out of all components of an optical computer, processing unit is the most desired.
As there are many nonlinear optical effects, many different possibilities have been tried in order to
create an optical logic gate. They usually rely on Kerr effect. Kerr effect is a quadratic electro-optical
effect, where material alters its refraction coefficient when electric field is applied.
(1) Where λ is the wavelength of the light, K is is the Kerr constant and E is electric field.
This principle is used to modulate signals for communication, as response of the material is very
swift. Electric field could also be a consequence of another beam, if the beam intensity is high
enough, making the process a third order nonlinear effect. Third order optical nonlinearity means
polarization depends on the cube of electric field,
(2)
Such transistors rely on cross phase modulation, where phase of one beam is influenced by another.
{
} (3)
is the wavelength of light in free space, l length of the material where beams interacts, is
second field and
is third order nonlinear tensor, where only part responsible for cross phase
modulation is taken.
Response to the light is swift; therefore their possible operating frequency is high, much higher than
of current electronic ones. However, most materials have a very small nonlinear index. For a
reasonably small transistor (to compete with electronic ones in terms of frequency/size), a very high
electric field is needed, meaning powerful lasers are required. Optical transistor major problem is
also material inability to sustain laser power without damage. When frequency of pulses gets high
enough to compete with silicon transistors, optical ones incinerate, limiting their maximum operating
frequency to less than 1 Mhz. Either a material with very high nonlinearity needs to be found or a
material needs to sustain very high power of lasers in order to increase this limit. First is preferred, as
a very powerful laser would be also impractical for computing. Wall Street Journal (Jan. 30, 1990)
labeled such a material "unobtainium,” as throughout all the years and funding, nothing was found.
Situation has not improved much since then.
Another material optical transistors could be based on are photonic crystals, or photonic bandgap
crystals they are also called. These crystals are made of a material with periodically varying refraction
index. While 1D photonic crystals have been known for over a century and frequently appear in
nature, first 2D crystal for optical wavelength has been fabricated in 1996. Their major problem is still
fabrication, as there is no known efficient method for creating such an array in 3D without defects.
Currently, majority of research on this field is focused on producing the crystals, mainly through self-
assembly. [7]
For computation using photonic crystals, nonlinear defects in crystals are required. Even a simple
scheme, where only 3 nonlinear rods are inserted in a 2D photonic crystal, exhibits bistable
transmission. This can be used to create a simple switching device, where one beam sets the path for
the second.
Image 4: Effects of nonlinear rods on transmission in a configuration shown in inset. Nonlinear rods are marked with black circles. [7]
Another phenomenon that showed promise and is being researched for optical computers is a
surface plasmon, a standing wave of free electrons on a metallic surface. When a photon hits metal,
electrons ripple with a specific frequency. These standing waves propagate through a metal, but
losses are very high. If such waves are confined only to the metallic surface and most energy travels
through a dielectric layer that is in contact with metal, losses are dramatically reduced and the signal
can travel much further even through a very narrow waveguide, below conventional limit. [8]
Although most research in this field is not useful for an all-optical computer at this stage, plasmons
are very promising for creating small lasers and waveguides. Both were demonstrated, but major
advances are still needed in order to be of practical use.
Optical logic gates also have the problem they need a lot of signal regeneration after gate. As optical
signal regeneration is difficult, some predict that optical computing and possibly even routing is
dead, or at least is not progressing the way scientists hoped for. [6]
On the other hand, some are more optimistic: ““For the last five years or so it has been possible to
build an optical computer chip, but with all-optical components it would have to measure something
like half a meter by half a meter and would consume enormous power. With plasmonics, we can
make the circuitry small enough to fit in a normal PC while maintaining optical speeds,” explains
Anatoly Zayats , a researcher at The Queen's University of Belfast in the United Kingdom.” - Quote
from [9].
Current status: Such a computer is possible, but not feasible. Many significant improvements are
required in order to compete with electronic processors for computers.
5. Storage Write once, read many (WORM) optical storage has been around for quite a long time, since the first
CDs in 80s, which even enabled more storage density than magnetic disks at the time. CD was
followed by a DVD, this one by Blu-ray. After Blu-ray, holographic storage is predicted as the next
stop. First stop for holographic storage is to create a write once disk, same as with other types of
optical storage. But, in order to make a proper computer storage using this technology, it is required
that a disk can also be rewritten multiple times. Rewritable holographic storage could be based on
photorefractive effect, an effect where a material alters its refractive index when exposed to light. If
a material exhibits Pockels effect, change of refraction coefficient is
(4)
Where n is a refraction coefficient without applied electric field, r is an electro-optic coefficient and E
is the electric field.
For the effect to be applied to holograms, process works in several steps. First, interference between
reference and signal beams creates a pattern of light and dark fringes throughout the crystal. In
regions with bright fringes, electrons absorb photons and promote into conductive band. Electrons
diffuse around the crystal (or drift due to photovoltaic effect), while holes must remain stationary.
Electrons may, with some probability, recombine with holes or fall back to the traps made by
impurities, where they cannot move. With more electrons in dark areas and holes in bright, there is a
net electric field, called space charge field, which causes, via electro-optic effect, refractive index to
change, creating a grating. Grating reflects light, recreating the original signal beam pattern, when
reference beam strikes the grating at the same angle.
The following setup does enable multiple writes, but every read would excite some of the trapped
electrons and decrease strength of the hologram. After several reads, image would not be
recognizable anymore.
The idea would need to be realized using two-photonic write in order to be practical. Two-photonic
setup uses 2 different dopants in order to create shallow and deep traps. When writing an image, a
pulse of short wavelength excites electrons to conduction band. They quickly fall back to a shallow
trap. From shallow traps, signal and reference beam can excite them back in conduction band, where
they diffuse like before. Afterwards, electrons fall back, first to a shallow trap, then to a deep. When
in deep traps, reference beam does not have enough energy to excite them back to conduction band,
greatly increasing durability of the hologram.
Image 5: Two-photonic write, described above. [10]
Additionally, storage could be written on in the Fourier space, meaning a single point of data does
not map to a single part of crystal, but rather to the whole, thus decreasing defects due to
imperfections.
Using all these techniques, holographic rewritable storage could at least compete with magnetic
disks with respect of reliability. Polymeric holographic storage, on the other hand, is among the most
durable WORM storages and could compete even with magnetic tape in terms of reliability.
There is also a possibility to create several holograms on the same space using different angle of
reference beam, or some other way of multiplexing signal (discussed in [10]). This increases storage
density, but decreases speed as reflection is weaker. Practical upper limit is around one hundred
holograms in a material 1 mm thick.
As the capacity of magnetic disks increase, holographic storage seems unable to bring any significant
advantages. First announced holographic disk would offer only 300 GB of capacity with up to 1.6 TB
to follow. Magnetic disks offer more capacity now than proposed storage could in the future. But
there is one major possible benefit of holographic storage – speed. Holographic drive enables reading
and writing whole set of data in parallel, enabling very high read and write speed. On the other hand,
all current data is written sequentially.
Due to lack of any moving parts, access time would be very low as well compared to magnetic drives,
where needle needs to move. Announced products, on the other hand, would be first severely
limited by write speed of only 20MB/s, which was said to increase later on, to be comparable with
current disks. [11]
The other possible benefit compared to current drives is longevity and reliability, both mostly valued
for permanent storage and not as useful for home computers as speed and capacity are. Magnetic
tape is currently dominant in this space with no real alternative.
Major difficulty of holographic storage and the reason it has not appeared yet is the material. The
material needs to have traps of correct energy, enable high write density and speed, would be
without defects that decrease performance and cheap enough to make, in order to be in commercial
products. Most researched material was Lithium Niobate, while many others have been tried in
laboratories, including proteins. Polymers are also currently used, but for permanent holograms.
One of the most noticeable companies, working on holographic storage, InPhase technologies, spun
off Bell Labs in 2000. The company made several claims when will they release their storage, first as
300GB disks with up to 1.6TB later, but nothing was released to the market. Recently, the company
had their assets seized, as they were unable to pay taxes. Similarly, Optware made several claims
about their upcoming products, while none actually appeared on the market in the end, and the
company does not exist anymore either.
General Electric and IBM continue research on holographic storage, but neither announced
upcoming products yet.
6. Memory Memory for systems seeks to have as high bandwidth as possible with lowest latency possible, while
having the largest size possible. These requests are mutually excluding, so compromises are made,
usually in favor of speed and latency. Optical competition to traditional memory seem very promising
as replacement, theoretically enabling high bandwidth and low latency, while being able to scale size
as well. Such optical system could be based on several effects. First possibility was discussed with
hard disks and is holographic data storage. Further option is a temporary storage of light pulses.
This storage relies on light to excite electrons in conduction band, generating pairs of electron-hole.
Electric field is applied to modulate potential and trap electrons and holes in spatially separate
potential minimums, limiting recombination. This energy can be later re-emitted in a short flash of
light, when potential is released. Due to the mechanism, only energy can be stored, coherence is not
preserved. The idea was shown to work with temperatures around 100K, light was stored for several
, while energy was released on demand. Further improvements could enable such storage for a
longer period of time and at a room temperature. As coherence cannot be preserved, this is not as
interesting storage possibility as the holographic one. [12]
In addition to these possibilities, light can be also slowed down or trapped in a resonator. Resonator
based storage would require mirrors with total reflectivity as there would be 300 reflections of
mirrors if we wanted to store light in 1 m long resonator for only 1 . Similarly, fiber based delay
requires low signal attenuation and ability to release stored pulses on demand at once. Both
solutions are possible, but more difficult to realize, compared to holographic storage. Another
possibility is to stop light and release it on demand.
Light stopping is a process in which light is slowed from vacuum speed at least two decades down.
However, for a practical memory cell, speed needs to be reduced to 0 on demand, otherwise
maximum storage time would be limited by memory length. Using a material with very high
refraction coefficient would also result in very high losses; therefore this is not a practical solution.
Possibility is to use Rabi oscillations, a quantum phenomenon where two states are coupled through
electric field.
Rabi oscillations Suppose we have a system with two eigenstate energies, ⟩ , ⟩ , and we shine a
light with frequency of We start perturbatively, where perturbative term of
hamiltonian is
[ ⟩⟨ ⟩⟨ ] (5)
New wave function is a linear superposition of our ground states.
| ⟩ ⟩ ⟩ (6)
Inserting wave function in Schrödinger equation, we get relations between coefficients A and B:
(
)
(7)
Probability of excited state therefore varies with a typical frequency, called Rabi frequency.
(8)
Frequency can also be written in generalized form, when field frequency is not equal to energy
difference,
√
| | (9)
Intuitively, light behaves as if the matter was periodically absorbing and then re-emitting photons
due to stimulated emission. Each such cycle is called a Rabi cycle. [14]
If we calculate eigenstates for the new Hamiltonian, we see that states split due to polarization,
energy difference between new states equals . These new states are called dressed states.
Therefore, matter is more transparent to the light it would be without state splitting.
We can use the effect with 3 states in configuration, where 1 wave, called coupling, couples states
2 and 3, while second (called probe) couples 1 and 3. As states 2 and 3 require much higher energy,
they are not occupied at start. First, coupling wave is activated, but as states 2 and 3 are empty,
there is no oscillation. Probe wave is then activated adiabatically. Due to probe wave, state 3 splits.
Oscillations interfere, resulting in impossibility to have any particles in state 3. Therefore, absorption
is not only a superposition of 2 absorption lines, but its imaginary part does fall to 0 in the middle,
while real part of the refraction index has a very large derivative.
Image 6: Real and imaginary part of refraction index without coupling field (dashed line) and with coupling field (solid line) [13]
If we deactivate coupling wave with light in material, energy is stored in state 2, meaning
propagation of light stops, till coupling wave is reactivated, when the light resumes propagation. As
there is spontaneous state transition through emission, strength of the signal decays exponentially.
Currently, lowest speed in a viable material was 80
, achieved in ruby vapour at 360K. Lower speeds
and even total stop were achieved, but vapour needed to be cooled to several hundred nK, way
below practical usage for storage. Light was also stopped only for some . Using current memory
production techniques and materials, light was stopped to 1/100 c. Therefore, it is possible to slow or
even stop light using this process, but it is not currently useful for memory.
Out of all mentioned technologies, only holographic seems achievable in not too distant future, while
none are actually close to commercial production.
7. Conclusion Optical computing is, at the moment, possible, but not practical. For servers, process of migration
towards optical connections even on a smaller scale has already started, while other system
components are currently not being focused on. For personal computers, situation is even less
promising, as technology is driven by its cost, not absolute performance for any price. Estimations
when all-optical computers could appear are currently only guesses, as technology is not ready even
for prototype computers, let alone massive commercial production.
8. References [1]: www.news.cornell.edu/stories/Feb08/4waveregen.ws.html; 21.3.2010
[2]: http://www.fujitsu.com/global/news/pr/archives/month/2005/20050304-01.html; 21.3.2010
[3]: M. Takenaka, K. Takeda, Y. Kanema, M. Raburn, T. Miyahara, H. Uetsuka, Y. Nakano, 320Gb/s
optical packet switching using all-optical signal processing by an MMI-BLD optical flip-flop, Proc. 32nd
Eur. Conf. Opt. Commun. (ECOC 2006), TH4.5.2, Cannes, France, 2006.
[4]: http://www.research.ibm.com/photonics/publications/ecoc_tutorial_2008.pdf 21.3.2010
[5]: Board-to-Board Optical Interconnection System Using Optical Slots; In-Kui Cho et al., IEEE
Photonics Technology Letters, Vol. 16, no. 7, July 2004
[6]: All-Optical Computing and All-Optical Networks are Dead; Charles Beeler, El Dorado Ventures and
Craig Partridge, BBN
[7]: http://braungroup.beckman.illinois.edu/photonic.html, 21.3.2010
[8]: Lasers go nano, Francisco J. Garcia-Vidal and Esteban Moreno, NATURE|Vol 461|1 October 2009
[9]: http://www.alphagalileo.org/ViewItem.aspx?ItemId=61986&CultureCode=en 21.3.2010
[10]: Holografsko shranjevanje podatkov, Blaz Kavcic, Ljubljana, November 2007
[11]: http://www.inphase-technologies.com/, 21.3.2010
[12]: Semiconductor based photonic memory cell, S. Zimmermann, et al., Science 283, 1292 (1999)
[13]: Pocasna svetloba, Martin Strojnik, May 2008
[14]: “Laser Electronics”, J. Verdeyen, 3rd ed., Chapt. 14.