スライド 1

表
PPT Created:
Alvaro
28/10/2002
題
機関名・研究者名（研究グループ名）
Introduction
A lot of effort has been put on free-space interconnects as a way to alleviate the wiring congestion inherent to planar VLSI technology. However, industry is
still awaiting for a reliable support, comparable in robustness to optical fibers (1D) or planar guided-wave optics (2D). We present here first steps towards a set of
robust, pre-aligned 3D fiber-based interconnection blocks, usable as elemental building modules for most parallel computing architectures. Thanks to
the guided nature of the interconnect, we can expect these to achieve better efficiency than diffractive optics, while the use of a third dimension will certainly
eliminate cross-talk.
Also, multistage architectures using regular plane-to-plane wave-guide based
interconnections may represent (today) the most advantageous hardware
implementation of a theoretically-optimal architecture for massively interconnected
systems, so we have concentrated on the design of inter-stage single or multiplepermutation modules.
Data
flow
Interconnection
module
Photodetector
array
Research
Interconnection
module
…
Interconnection
module
VCSEL
array
Elementary
Processor Array
Although technologically challenging, there is no fundamental reason excluding
the use of optical fibers as a way to implement complex interconnection patterns
between plane arrays. However, compared to free-space optics, alignment (of
both inputs and outputs) may represent a real challenge. Our solution consist of
using pre-aligned micromachined fiber-holder arrays.
Scalability and modularity of the
blocks is another important issue
being explored through fractal and
row/column recursive
decomposition.
Results
Active alignment is being explored as a way to deal with the
interconnection difficulties inherent to the cascading of the modules, or
even for selecting (switching) them.
VCSEL
Interconnection
optics
VCSEL
Vibration
generator
Vibration
generator
Quadrant
photodetector
Interconnection
optics
Quadrant detector
Controller
xy Piezo Stage
Piezo microstage
Control signal
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
00
Misalignment Reduction Ratio
(Active Alignment ON / OFF)
20
40
60
80
100
Frequency [Hz]
120
140
Time Division Multiplexed Banyan network with global switching
module control
module control
output
module control
{ (1) , i}
bi-module
PE array
PE array
A different, interesting issue worth to consider is the use of modules containing
several different, independent addressable permutations (sharing the same
folding properties). This can lead to a concurrent multistage interconnection
paradigm not based on size-reduced (elemental) crossbars (as in most multistage
networks) but on cascaded permutation-reduced crossbars instead. Promising
buffer-based architectures based on this paradigm as well as routing algorithms
are currently being explored in our lab.
input
Further research
Active Alignment
ON
We plan to use several
similar modules to build
a spanned hypercube
network with TDM
(Time-DivisionMultiplexed) weak
communication links.
Active Alignment
OFF
Misalignment measurement
{ (2), i}
bi-module
PE array
As a demonstration of these principles, we implementing whole shuffleexchange multi-stage interconnection networks (using these modules and the
optoelectronic processor arrays developed in our lab for the Vision-Chip project).
We have designed, fabricate and tested a folded 16-input/output exchange
permutation module, where cross-talk between channels is made negligible.
{ (3), i}
bi-module
Publications and patents
・特許出願件数：
・投稿論文数：
・学会発表
：
件: one patent pending.
件: 2 published, 2 writing
件: 3 presentation (2 in international conferences)
大阪府地域結集型共同研究事業

Download Report