lindsay@gandalf.cs.cmu.edu (Donald Lindsay) (04/09/91)
I posted recently about using lasers to communicate short distances,
e.g between chips. I was asked by mail about how you route light, so,
he's some answers. Disclaimer: I don't do this stuff: corrections and
donations welcome.
The most obvious idea is to use optic fibers, because you can buy
them, and they have unbelievable optical quality. This clearly solves
the routing problem - it would be a bit like wire-wrap. But how to
attach a fiber to a chip?
There are several answers. I think one group made a notch in the edge
of a chip, so that a fiber in the notch would be properly aligned
with an edge-emitting diode. Now that we can make arrays of
face-emitting lasers, there's a simpler answer: just epoxy the fiber
end onto the chip. One of the lasers will happen to be on the fiber's
optical axis, and one is enough. One Bell Labs estimate was that they
could build 1024 lasers in the space of a single conventional bonding
pad.
Fibers are bulkier than we might like, and aren't integral, so other
ideas have been followed. One of the older ideas is optical layers.
Various people have sent light sideways through films [as in the
phrase "thick film", not as in camera film]. It turns out that e.g.
lenses can be formed by manipulating the fabrication of the film.
Since beams of light can cross and interpenetrate without
interference or heavy S/N problems, there were suggestions for e.g.
shuffle-exchange networks sharing the same film.
There has been talk about "free space" communication, meaning that
beam travel isn't guided. There problem here is to keep the
geometries fixed well enough so that multiple beams can be used,
without problems from e.g. vibration.
The "free space" doesn't have to filled with air, however. It could
be filled with the next wafer in a stack of wafers. In this idea, you
send light vertically, right through a wafer. Since a wafer is only a
few mils thick, it doesn't necessarily matter that it isn't all that
transparent. One can even do interesting S/N tricks, like etch
Fresnel lenses onto the back face of the next wafer up.
A newer "free space" proposal fills the free space with quartz.
Specifically, Bell Lab's "optical LEGO(tm)" would use rectangular
prisms of quartz, with all but one face silvered. The remaining
face is etched in various useful ways. They have fiddled with
making gratings and lenses: the optical path is essentially
^
\ /
-----------
| \/\/\/ |
-----------
with the lenses (etc) in the middle region of the top face. The
suggestion is to place chips there, face down, and run the optical
interconnection through the quartz. Further, it is suggested that
multiple quartz blocks be connected, LEGO fashion, ie
+---------+
| |
+------+-+-----+-+------+
| | | |
+--------+ +--------+
The idea here is that by etching mechanical alignment guides, it
would be possible to plug the pieces together (while clean!) with
geometrical accuracies fairly near the order of lithographic
accuracies.
To sum up, I don't know that any of the above will become popular.
But these ideas, and the numerous non-optical proposals, have
completely convinced me that little gold wires have got to go. The
bonding pads are football fields: the power is brutal: pin
inductance chokes the signal. The physics to to better is there: a
mere few billion dollars of engineering development costs shouldn't
stop an industry like ours.
--
Don D.C.Lindsay .. temporarily at Carnegie Mellon Roboticspeng@hpsciz.sc.hp.com (Peng Lee) (04/10/91)
Here's one more idea for you:
Laser
| | |
| | |
| | |
\ / \ / \ /
------------------- Chip A
------------------- Chip B
As long at the laser wavelength (IR) is less than the Si 's band gap,
the silicon will be transparent to the Laser.
On chip A, one can use the liquid crystal to modulate the light. On
chip B, one can use the CCD to detect the IR (a very bad idea, someone please
point out a different way to do it.).
The only non-optical interconnection for this device is the power and
ground for each chip.
While I was in school, I used to fatasize an one cubic inch
workstation. It is a stack of chips with lasers on top and bottom as optical
sources. This workstation data and address buses are optical interconnected
with as much as 1 GByte/sec transfer rate even though the LC switch is
only toggling at 10MHz. An one cubic inch Kille-Nano.
Let me know if anyone likes to build something like this.
-Pengdavidsen@crdos1.crd.ge.COM (Wm E Davidsen Jr) (04/11/91)
In article <510001@hpsciz.sc.hp.com> peng@hpsciz.sc.hp.com (Peng Lee) writes: | On chip A, one can use the liquid crystal to modulate the light. On | chip B, one can use the CCD to detect the IR (a very bad idea, someone please | point out a different way to do it.). Why would you do this rather than use an IR LED on the sender chip, avoiding having to pipe the light through multiple chips? With only a few levels of ship there would be a power disipation saving, but with the cube you describe later in your posting I would think the loss of realestate to windows would be worse than the heat problem. Consider this a question rather than a criticism, I can't even ballpark the tradeoffs in my head, so I don't imply in any way you're wrong. -- bill davidsen (davidsen@crdos1.crd.GE.COM -or- uunet!crdgw1!crdos1!davidsen) "Most of the VAX instructions are in microcode, but halt and no-op are in hardware for efficiency"
bruceh@sgi.com (Bruce R. Holloway) (04/12/91)
In article <510001@hpsciz.sc.hp.com> peng@hpsciz.sc.hp.com (Peng Lee) writes: > > The only non-optical interconnection for this device is the power and >ground for each chip. > > While I was in school, I used to fatasize an one cubic inch >workstation. It is a stack of chips with lasers on top and bottom as optical >sources. This workstation data and address buses are optical interconnected >with as much as 1 GByte/sec transfer rate even though the LC switch is >only toggling at 10MHz. An one cubic inch Kille-Nano. > > Let me know if anyone likes to build something like this. Yeah, I'd like to build one. I'd want at least one solar cell & color LCD on the sides. Then I could take it to the beach and look at the pictures. Regards, bruceh
sysmgr@KING.ENG.UMD.EDU (Doug Mohney) (04/12/91)
In article <1991Apr11.213418.26877@odin.corp.sgi.com>, bruceh@sgi.com (Bruce R. Holloway) writes: >> >> While I was in school, I used to fatasize an one cubic inch >>workstation. It is a stack of chips with lasers on top and bottom as optical >>sources..... . An one cubic inch Kille-Nano. >> >> Let me know if anyone likes to build something like this. > >Yeah, I'd like to build one. I'd want at least one solar cell & color LCD >on the sides. Then I could take it to the beach and look at the pictures. No! No! "Top" covered with a solar cell, one side with color LCD, one side with CCD so you could TAKE pictures (Yah, who said technology couldn't have a practical use). Signature envy: quality of some people to put 24+ lines in their .sigs -- > SYSMGR@CADLAB.ENG.UMD.EDU < --
mpope@ATT.COM (michael.t.pope) (04/13/91)
[submitted on behalf of the author] One advantage of free space optical techniques over fiber is that you're not creating a "wire" for each connection. Instead you're generating an array of "spots", each only a couple of microns across and then transferring that as an image "en mass" to a destination receiver array. Bulk optical elements such as beam splitters, lenses and mirrors do the directing. The "LEGO Block" idea was proposed by Jurgen Jahns and Alan Huang in the Digital Optics Research department here at Bell Labs [1]. The lenses and mirrors are created by etching diffractive optical elements on the surface of a glass substrate. The light propagates inside the substrate as if it were free space, reflecting off the parallel surfaces and the optical elements etched on them. This is nice because everything is lithographically defined, giving great accuracy and ease of manufacture. LCD modulators are cheap, but have a problem for high speed applications: they can only modulate at the rate of 100's of KHz. We are using Self Electro-Optic Effect Devices (SEEDs) as optical modulators, or Surface Emitting Microlasers (SELs) as directly modulated light sources. Receivers can be made out of regular digital CMOS if you're careful. Speeds of 100's MHz to GHz are possible with these devices, and beam spacings are around 20 microns. Using optical wavelengths that make silicon transparent is appealing, but (amongst other things) detection is hard. We're working towards manufacturable electro-optical systems where the regular, high density, high speed connections are made optically through the substrate, and slower, less regular connections are made with wires on the surface of the substrate. One might imagine for example a 1024 channel system bus connecting cache and main memory - this could reduce the miss penalty, allowing the use of very wide cache blocks. One of the main problems remains the cost-effective integration of GaAs based optical sources/modulators with silicon technology. I'd be happy to provide further references. [1] J. Jahns and A. Huang, "Planar integration of free-space optical components," Applied Optics, vol 28, page 1602, May 1989. Alex Dickinson alex@vax135.att.com Digital Optics Research AT&T Bell Labs, Holmdel, NJ
mmm@cup.portal.com (Mark Robert Thorson) (04/14/91)
Pardon me if I missed somebody else mentioning it, but Forrest Mims published an article on optoelectronics in Popular Electronics back about 20 years ago, in which he described an ingenious method of cheaply attaching optic fibers to chips. He takes an LED and drives a soldering iron straight through the lens till he just touches the chip. Then he puts a drop of glue in the hole and pushes in the optic fiber. He notes that LED's have a lot of photosensitivity, and can be used as both transmitters and receivers.
gd@geovision.gvc.com (Gord Deinstadt) (04/15/91)
davidsen@crdos1.crd.ge.COM (Wm E Davidsen Jr) writes: >In article <510001@hpsciz.sc.hp.com> peng@hpsciz.sc.hp.com (Peng Lee) [suggests using external lasers perpendicular to a stack of chips for interconnect:] > Why would you do this rather than use an IR LED on the sender chip, >avoiding having to pipe the light through multiple chips? I'm not sure about LEDs, but I know it's hard to make a laser that emits perpendicular to the surface of the chip - most are edge-emitters. But you could put mirrors or light guides along one edge of the chips. Certainly the gain in speed vs LCD switches would be worthwhile. If you really need the 2-d interconnect, you could also interleave the chips with electroluminescent emitters, and perhaps use lithium niobate (or some such) modulators for higher speed. -- Gord Deinstadt gdeinstadt@geovision.UUCP