charlie@sunoptics.caltech.edu (Charles Stirk) (07/24/90)
Dear Andy, Sorry I don't know how to post to usenet, so if you could put this there for me I would appreciate it. The latency of an optical interconnect, like an electronic one, depends as much on the circuit as on the individual device characteristics. Optoelectronic devices are composed of diodes, transistors, resistors inductors and capacitors. Only with very exotic devices should the device characteristics be very much different than those found in electronics. Since the maturity of the optoelectronic devices lags behind those of electronics in most cases, due to a smaller and younger market, it is difficult but not impossible to make a case for optical interconnects in computer architecture purely on device characteristics. The lower noise and higher bandwidth that is an optical circuit advantage in telecommunications has also been useful in some computer architecture applications like buses and peripheral connections. The advantages come from the passive part of the circuit, the optical waveguide. The advantages gained by better circuit layout with optics can be significant and are easy to demonstrate in principle. For instance, in vlsi complexity theory the minimum amount of chip area it takes to layout the shuffle exchange graph, which is useful in parallel processors, is at least (N/logN)^2, where N is the number of nodes in the graph. My collegues and I demonstrated an optical shuffle exchange network that uses only N units of area. Area costs money in vlsi yield models and small area circuits can also mean high speed. Similar arguments can be made for other graphs like the hypercube. Practical demonstrations of these advantages, however, are difficult due to the technological immaturity alluded to above. That doesn't mean demonstrations haven't been attempted, they just don't irrefutably show a significant practical advantage. A good example of this is public relations extravaganza coming out of Bell Labs in the last year on optical computing. In addition, the amount of freedom available in choosing values for the parameters for the devices and circuits in a comparison between optics and electronics is wide enough so that almost any conclusion can be drawn, both for and against optics. It is an active area of research, nonetheless. For an introduction to the field of optical interconnects, I recommend the papers by J. W. Goodman of Stanford and his associates published in Proc. IEEE, Applied Optics and Optics Communications since 1984. Sincerely, Charlie Stirk charlie@sunoptics.caltech.edu -- Andy Glew, andy-glew@uiuc.edu Propaganda: UIUC runs the "ph" nameserver in conjunction with email. You can reach me at many reasonable combinations of my name and nicknames, including: andrew-forsyth-glew@uiuc.edu andy-glew@uiuc.edu sticky-glue@uiuc.edu and a few others. "ph" is a very nice thing which more USEnet sites should use. There is an info-ph mailing list, or, contact Steve-Dorner@uiuc.edu. UIUC has ph wired into email and whois.
aglew@oberon.crhc.uiuc.edu (Andy Glew) (07/24/90)
A while back I posted about an electro-optic bus to comp.arch. It generated a flurry of responses, but nobody provided the real data that one correspondent rightly suggested would be critical to an electro-optic bus. I am hoping that sci.electronics or comp.lsi or sci.physics readers may be able to fill in the gaps. (For these newsgroups the original post is included at the end of this post). We know that optics can achive high bandwidth, ie. throughput. But what is the latency of an electro-optic system? Eg. what would be the time to covert from a - GaAs electronic (voltage) signal, - to optics (typically via a laser diode, hopefully on the GaAs chip) and then to receive - native Si optical receiver. - amplify the signal to CMOS levels. For good measure add the time to drive a wire between the Si and GaAs chip. What is the ballpark for this electrical->optic->electrical conversion? Tens of nanoseconds? Sub-nanosecond? Any experience people have in timing discrete electro-optical systems would be helpful, providing at least an upper bound. The original post: >Newsgroups: comp.arch >From: aglew@oberon.crhc.uiuc.edu (Andy Glew) >Subject: Electro-optic bus >Organization: University of Illinois, Computer Systems Group >Distribution: comp >Date: 16 Jul 90 21:59:01 > >Here's a throwaway - the electro-optic star bus: > > We all know that electrical interconnects such as busses have >major performance problems - loading, skews, etc. [*]. >The longer and the more things you talk to, the slower. > > We all know that optics has much more bandwidth than electrical >busses - right? (Maybe I should put a smiley here?) > > We all know that the biggest problem with optics is that Si >doesn't interface to it well. GaAs does, but GaAs doesn't achieve >Si's levels of integration [**]. > > What is, perhaps, less well known is that Si can build fairly good >*receivers* for optics; Si just cannot build good transmitters. Even >MOSIS's CMOS processes can build fairly good receivers. > > > Throw-away idea: take, say, 4 Si microprocessors. Give them each a >point-to-point (easier to make fast) electrical interconnect (wires), >from the Si chips, to a GaAs chip. Let the GaAs chip take these 4 >sets of electrical signals, and compress them all onto a faster >optical bus that is sent back to the 4 microprocessors. > Ie. use the GaAs chip as the hub of a star, with incoming signals >in electronics, and outgoing signals in optics. > The Si chips all receive the optical signals with native Si >receivers. > Since all of the signals are broadcast on the optics, we >effectively have a bus. If the Si receivers and logic can be made fast >enough, then we could snoop and do all those sorts things that are >nice to do on busses. Ie. bus bandwidth is no longer the constraint - >the constraint is how fast the Si receivers and/or the Si cache >snooping logic can be made. > > > > >Waiting anxiously to be shot down... > > >[*] this is not to say that there aren't people who believe that > existing bus technology cannot be milked. Such as making existing > 25 MB/s busses run at more than 4 MB/s under real usage patterns. > But that's another story. > >[**] I am, of course, neglecting the problem of coupling optical fibers > (if they were fibers) to the Si chips. >-- >Andy Glew, aglew@uiuc.edu -- Andy Glew, andy-glew@uiuc.edu Propaganda: UIUC runs the "ph" nameserver in conjunction with email. You can reach me at many reasonable combinations of my name and nicknames, including: andrew-forsyth-glew@uiuc.edu andy-glew@uiuc.edu sticky-glue@uiuc.edu and a few others. "ph" is a very nice thing which more USEnet sites should use. There is an info-ph mailing list, or, contact Steve-Dorner@uiuc.edu.
loving@lanai.cs.ucla.edu (Mike Loving) (07/24/90)
In article <AGLEW.90Jul23203937@oberon.crhc.uiuc.edu> aglew@oberon.crhc.uiuc.edu (Andy Glew) writes: >A while back I posted about an electro-optic bus to comp.arch. It >generated a flurry of responses, but nobody provided the real data >that one correspondent rightly suggested would be critical to an >electro-optic bus. > I am hoping that sci.electronics or comp.lsi or sci.physics >readers may be able to fill in the gaps. (For these newsgroups the >original post is included at the end of this post). [some stuff removed] > >Any experience people have in timing discrete electro-optical systems >would be helpful, providing at least an upper bound. A while back I worked with Dr. Norm Matloff at UC Davis on optical data transmission in computers. It differed in one major respect from your scenario, no lasers on the chips. Instead a thin electo-optic film is coated onto the chip. This film will polarize light in different ways depending on the voltage applied across the film. By then running the light through polarizing filters, one gets intensity modulation. You then run the light into what is basically a sense amp with optical reception (reverse bias diode or photo transistor). In the stuff we were doing, with the process parameters we had, we could distinguish between ones and zeros reliably if we had greater than 1% difference in incoming intensities. This took about 10 nanoseconds, but in theory, you transfer the entire contents of a chip to another chip this way, you just have focus and alignment difficulties that need to be worked out. When I left there 2 years ago, the electro-optic films were still not quite good enough to do the polarization job. norm matloff = matloff@iris.ucdavis.edu for more info than I've got. Mike Loving loving@lanai.cs.ucla.edu
charlie@sunoptics-gw.caltech.edu (Charles Stirk) (07/25/90)
Inductance can be very important in packaging. The reason for this is common ground bounce, where large current transients through power and ground pins due to large numbers of devices switching rapidly and the inductance on the pins of the package cause the power rail to drop in voltage and the ground rail to come up. There are ways to reduce this effect. The first is large numbers of power and ground pins for high-speed and high-current circuits. The second is using large decoupling capacitors in the package to keep the power and ground levels relatively constant. The third method is to design circuits that don't have current transients through the pins. Another, though much less severe inductance problem is the crosstalk noise due to induced currents in pins and wires. Designers get around this by running a ground pins between the signal pins. I did not see the statement that optical fibers are limited to about 1Gbit/sec, so I'm not sure if the author meant present state of technology or fundamental limits. Optical fibers have supported bandwidths far in excess of 1 GHz, and the fundamental limit is ridiculously high. The main limit on increasing bandwidth at a fixed bit error rate is the power budget between the transmitter and receiver and channel crosstalk noise. Using wavelength division multiplexing of many channels in one fiber, extremely high-bandwidths have been demonstrated. In practice, one is limited more by the source and detector characteristics than the characteristics of the fiber. The time it takes to do electrical->optic->electrical conversion is limited by the acceptable bit error rate, the optical power of the transmitter, the sensitivity of the detector and amplifier, the noise sources, and the time of flight. All of these factors must be considered when determining the latency of a link. Increasing the first two and decreasing the second two will decrease the latency of the link. The time of flight is significant because light travels at about a foot per nanosecond. A paper that discusses several of these issues and their effect on bandwidth is by Anil Jain in Optical Engineering a couple of years ago. A note of caution when designing optical interconnects into computer architecture. Present architectures have been designed with the the known characteristics of electrical interconnects in mind. Optical circuits, however, have a markedly differents set of characteristics. While some are improvements over those in electronics like bandwidth, distance and noise, and can be used to incrementally upgrade the present architectures, the cost may not be worth it. Many optical devices are presently much more bulky, expensive and unreliable than their electronic counterparts. They will probably only be commercially feasible in high-end systems, where their added cost can be justified, until these factors change. Optical interconnects can have many disadvantages if not designed correctly. A true revolution in computer architecture could occur, however, when systems are designed from scratch to take advantage of the many useful characteristics of optical interconnects. Sincerely, Charlie Stirk charlie@sunoptics.caltech.edu -- Andy Glew, andy-glew@uiuc.edu Propaganda: UIUC runs the "ph" nameserver in conjunction with email. You can reach me at many reasonable combinations of my name and nicknames, including: andrew-forsyth-glew@uiuc.edu andy-glew@uiuc.edu sticky-glue@uiuc.edu and a few others. "ph" is a very nice thing which more USEnet sites should use. UIUC has ph wired into email and whois (-h garcon.cso.uiuc.edu). The nameserver and full documentation are available for anonymous ftp from uxc.cso.uiuc.edu, in the net/qi subdirectory.
aduncan@rhea.trl.oz.au (Allan Duncan) (07/27/90)
From article <AGLEW.90Jul23203937@oberon.crhc.uiuc.edu>, by aglew@oberon.crhc.uiuc.edu (Andy Glew): ... > We know that optics can achive high bandwidth, ie. throughput. > > But what is the latency of an electro-optic system? > Eg. what would be the time to covert from a > - GaAs electronic (voltage) signal, > - to optics (typically via a laser diode, hopefully on the GaAs chip) ^^^^^ > and then to receive ... A laser might not be the best - they have a delay from zero curent while a population inversion is created. Normally they are used with _varying_ intensity rather than off/on, so there is added complexity, as well as the (current) need to edge couple to the fibre. LEDs are better unless you want coherence. What you end up with is a pipeline, rather than a bus, maybe this will better match the current crop of pipelined CPUs :-) Overall, there is little, if any, advantage in electro-optic busses. If it was all optic, then that might be different, but even then you are only talking pipes, as the trip time to an adjacent slot is significant. Allan Duncan ACSnet a.duncan@trl.oz (03) 541 6708 ARPA a.duncan%trl.oz.au@uunet.uu.net UUCP {uunet,hplabs,ukc}!munnari!trl.oz.au!a.duncan Telecom Research Labs, PO Box 249, Clayton, Victoria, 3168, Australia.
berryh@udel.edu (John Berryhill) (07/28/90)
Furthermore, using lasers on an electro-optic bus is like using a chainsaw to cut butter. Not only is it possible to fabricate LED's that work up to 1 GHz, but they are MUCH more reliable. You use lasers when you need much higher bandwidth, higher power, and lower dispersion. Lasers ought to be used to multiplex lots of signals and send them long distances over a single fiber. When you are going to be using a bunch of fibers to send multiple channels over short distances, use LEDS instead (preferably the ones I've designed for this very application :-) ). BTW, don't let optical coupling efficiencies bother you all that much. Coupling efficiencies are typically measured with fibers longer than the distances that we're talking about here. There are plenty of modes that decay exponentially, but provide significant power over just a few feet. -- John Berryhill 143 King William, Newark DE 19711
aglew@oberon.crhc.uiuc.edu (Andy Glew) (07/31/90)
>A laser might not be the best - they have a delay from zero curent while >a population inversion is created. Normally they are used with _varying_ >intensity rather than off/on, so there is added complexity, as well as >the (current) need to edge couple to the fibre. LEDs are better unless >you want coherence. > >What you end up with is a pipeline, rather than a bus, maybe this will >better match the current crop of pipelined CPUs :-) > >Overall, there is little, if any, advantage in electro-optic busses. If >it was all optic, then that might be different, but even then you are >only talking pipes, as the trip time to an adjacent slot is significant. So, what is the ballpark delay? Laser and/or LED? Tens or hundreds of ns? (I don't mean to be antagonistic. It's just that I've heard several times "there is a delay", without anyone quoting even any order of magnitude numbers). -- Andy Glew, andy-glew@uiuc.edu Propaganda: UIUC runs the "ph" nameserver in conjunction with email. You can reach me at many reasonable combinations of my name and nicknames, including: andrew-forsyth-glew@uiuc.edu andy-glew@uiuc.edu sticky-glue@uiuc.edu and a few others. "ph" is a very nice thing which more USEnet sites should use. UIUC has ph wired into email and whois (-h garcon.cso.uiuc.edu). The nameserver and full documentation are available for anonymous ftp from uxc.cso.uiuc.edu, in the net/qi subdirectory.
aglew@oberon.crhc.uiuc.edu (Andy Glew) (07/31/90)
If it's anything, I'd seriously think about space optics. Since I'm talking about a bus, I could imagine several GaAs LEDs (okay, not lasers) pointing up, with something to diffuse the light and spread it around to all of the Si CPU chips with their receivers on their top surfaces. Receive power and packaging was the reason I didn't state this in my first posting, although I did hint heavily. -- Andy Glew, andy-glew@uiuc.edu Propaganda: UIUC runs the "ph" nameserver in conjunction with email. You can reach me at many reasonable combinations of my name and nicknames, including: andrew-forsyth-glew@uiuc.edu andy-glew@uiuc.edu sticky-glue@uiuc.edu and a few others. "ph" is a very nice thing which more USEnet sites should use. UIUC has ph wired into email and whois (-h garcon.cso.uiuc.edu). The nameserver and full documentation are available for anonymous ftp from uxc.cso.uiuc.edu, in the net/qi subdirectory.
arnief@tekgvs.LABS.TEK.COM (Arnie Frisch) (08/01/90)
In article <1965@trlluna.trl.oz> aduncan@rhea.trl.oz.au (Allan Duncan) writes: >From article <AGLEW.90Jul23203937@oberon.crhc.uiuc.edu>, by aglew@oberon.crhc.uiuc.edu (Andy Glew): > >... >Overall, there is little, if any, advantage in electro-optic busses. If >it was all optic, then that might be different, but even then you are >only talking pipes, as the trip time to an adjacent slot is significant. > There are two significant advantages to electro-optic busses: 1. No crosstalk. A significant factor in wide bus high speed applications. 2. Wide bandwidth as a function of the separation of the transmitter and receiver. Reduction of dispersive effects, such as skin effect. For these reasons, most computer manufacturers are looking at architectures including optical busses for advanced machines. Arnold Frisch Tektronix Laboratories
aduncan@rhea.trl.oz.au (Allan Duncan) (08/01/90)
From article <AGLEW.90Jul30150521@oberon.crhc.uiuc.edu>, by aglew@oberon.crhc.uiuc.edu (Andy Glew): >>A laser might not be the best - they have a delay from zero curent while > So, what is the ballpark delay? Laser and/or LED? > Tens or hundreds of ns? > > > (I don't mean to be antagonistic. It's just that I've heard several > times "there is a delay", without anyone quoting even any order of > magnitude numbers). Depending on area etc it can be up to 10 ns for lasers. LEDs are a little quicker. Allan Duncan ACSnet a.duncan@trl.oz (03) 541 6708 ARPA a.duncan%trl.oz.au@uunet.uu.net UUCP {uunet,hplabs,ukc}!munnari!trl.oz.au!a.duncan Telecom Research Labs, PO Box 249, Clayton, Victoria, 3168, Australia.
aglew@oberon.crhc.uiuc.edu (Andy Glew) (08/01/90)
>>>A laser might not be the best - they have a delay from zero curent... > >Depending on area etc it can be up to 10 ns for lasers. LEDs are a >little quicker. Thanks for providing the ballpark number! Now, is that just the time for the electro-optic conversion? Does anyone have any ideas what the latency for reception by a native Si optic->electronic receiver would be? Especially taking into account the time to amplify from whatever signal levels the optical receiver produces, to normal Si logic levels? Let me do some informal calculations (mainly to prompt people who know better to correct my (mis)estimates): Say it's 5ns for the electro-optic latency. Say another 5ns for the reception and amplification to CMOS levels. Add 1ns for transmission time in the optical medium Add another 4ns for the point-to-point electrical interconnect. That gives maybe a 15ns latency. Well, that's nothing great... you never go wrong building a low-latency interconnect, and electronics can do better than this. How many signals could be crammed onto this 15ns latency electro-optical star? -- Andy Glew, andy-glew@uiuc.edu Propaganda: UIUC runs the "ph" nameserver in conjunction with email. You can reach me at many reasonable combinations of my name and nicknames, including: andrew-forsyth-glew@uiuc.edu andy-glew@uiuc.edu sticky-glue@uiuc.edu and a few others. "ph" is a very nice thing which more USEnet sites should use. UIUC has ph wired into email and whois (-h garcon.cso.uiuc.edu). The nameserver and full documentation are available for anonymous ftp from uxc.cso.uiuc.edu, in the net/qi subdirectory.
berryh@udel.edu (John Berryhill) (08/01/90)
In article <7903@tekgvs.LABS.TEK.COM> arnief@tekgvs.LABS.TEK.COM (Arnie Frisch) writes: >There are two significant advantages to electro-optic busses: > > 1. No crosstalk. A significant factor in wide bus high speed > applications. ...and hence immunity to EMI. > 2. Wide bandwidth as a function of the separation of the > transmitter and receiver. Reduction of dispersive effects, > such as skin effect. 3. Freedom from two-dimensional constraints. Optical waveguides can be fabricated on a 2D surface and intersect one another with little signal leakage. Alternatively, as has been alluded to, optical signals can be routed holographically through free space. Ideally, you'd have holograms that can be modulated in some fashion in real time so that your entire interconnect topology can be altered to fit the computational task at hand (the electro-optic effect in Lithium Niobate does have a small time persistence). -- John Berryhill 143 King William, Newark DE 19711
mcdonald@aries.scs.uiuc.edu (Doug McDonald) (08/01/90)
In article <AGLEW.90Jul31211132@oberon.crhc.uiuc.edu> aglew@oberon.crhc.uiuc.edu (Andy Glew) writes: >>>>A laser might not be the best - they have a delay from zero curent... >> >>Depending on area etc it can be up to 10 ns for lasers. LEDs are a >>little quicker. > >Thanks for providing the ballpark number! > >Now, is that just the time for the electro-optic conversion? Does >anyone have any ideas what the latency for reception by a native Si >optic->electronic receiver would be? Especially taking into account >the time to amplify from whatever signal levels the optical receiver >produces, to normal Si logic levels? > >Let me do some informal calculations (mainly to prompt people who know >better to correct my (mis)estimates): > > Say it's 5ns for the electro-optic latency. > Say another 5ns for the reception and amplification to CMOS levels. > Add 1ns for transmission time in the optical medium > Add another 4ns for the point-to-point electrical interconnect. > >That gives maybe a 15ns latency. > >Well, that's nothing great... you never go wrong building a >low-latency interconnect, and electronics can do better than this. > >How many signals could be crammed onto this 15ns latency electro-optical star? >-- Well, the rise time for a cheap ($0.10 [sic]) silicon detector (1N914) at 1.06 microns is 75 psec. That is 0.075 nsec. I would guess the amplification to 1-volt levels could be done easily (by high speed circuit standards) in 150 psec. For money you can get it down into the 10s of picoseconds. The risetime of emitter diodes (and emitter lasers too, if you always run above threshold, which is not exactly fun) is probably in the same ballpark. Please don't quote different times for the risetime of a 1N914 - that is the measured value, measured by me with a 50 psec light pulse and Tektronix's fastest sampling scope. It differs from the normal electrical response time because inm this case the thing remains always back-biased. There are commercial diodes that are that fast into the visible. Doug McDonald
berryh@udel.edu (John Berryhill) (08/02/90)
Most of the questions asked in this thread are answered in one or more of the following. If you want a broad overview of the possibilities, read Goodman's paper. If you are interested in specific performance requirements and capabilities, Hartman's paper is a good place to start. Goodman, Joseph W. et al. "Optical Interconnection for VLSI Systems" Proceedings of the IEEE, Vol. 62, No. 2 Fried, Jeffrey A. "Optical I/O for High Speed CMOS Systems" Optical Engineering, Vol. 25, No. 10 Hartman, Davis H., "Digital High Speed Interconnects: A Study of the Optical Alternative," Optical Engineering, Vol. 25, No. 10. Haugen, P.R. et al. "Optical Interconnects for High Speed Computing" Optical Engineering, Vol. 25, No. 10 Prucnal, Paul R., Eric R. Fossum, and Richard M. Osgood, "Integrated Fiber-Optic Coupler for VLSI Interconnects," Optics Letters, Vol. 11, No. 2. Wada, H, "Optoelectronic Integration Based on GaAs Material" Optical and Quantum Electronics, Vol. 20, p441 Dawson, L.R. et al. "Reliable, High-Speed LEDs for Short-Haul Optical Data Links" Bell System Technical Journal, Vol 59., No. 2. -- John Berryhill 143 King William, Newark DE 19711
hsu@sp26.csrd.uiuc.edu (William Tsun-Yuk Hsu) (08/02/90)
Thanks to John Berryhill for posting references. I checked and the Goodman article ("Optical Interconnections for VLSI Systems" by Goodman, Leonberger, Kung, and Athale) is actually in Proceedings of the IEEE, Vol.72 No.7. Bill Hsu
fmgst@unix.cis.pitt.edu (Filip Gieszczykiewicz) (08/02/90)
Greetings. On "Science and Technology", a TV (cable?) show: "company" has succesfully build the first optical "circuit". It's composed of 128 optical "gates". It's promising to produce a fully optical computer capable of near-speed-of-light operation by the year 2000...." Not sure of the source, and the "company" might have been AT&T, Bell Labs... something like that. ******************** Question1. Anyone hear of this, or better yet, worked on it? Question2. Feasible? For example how do you make an optical inverter? (I thought of a light-sensitive liquid that would react to light as does LCD panel to electricity...but that's not it...) Anyone? Take care. -- _______________________________________________________________________________ "The Force will be with you, always." It _is_ with me and has been for 10 years Filip Gieszczykiewicz "A man without a dream is like a fish without water." FMGST@PITTVMS or fmgst@unix.cis.pitt.edu "My ideas. ALL MINE!!"
olsen@artemis.ll.mit.edu (Jim Olsen) (08/02/90)
John Berryhill's list of references is a good one, but if you're looking for a quick overview, I'd suggest Hutcheson, L.D. et al., "Optical Interconnects replace hardwire", IEEE Spectrum, March 1987, pp. 30-35. To answer Andy Glew's question about latency, the Fried paper (Opt. Eng., Oct. 1986) gives a quick comparison of electrical and optical interconnect, and quotes 2-3 ns for electrical vs. 3-4 ns latency for optical interconnect for some 'typical' interconnects at 1 GHz.
arnief@tekgvs.LABS.TEK.COM (Arnie Frisch) (08/02/90)
In article <AGLEW.90Jul30150521@oberon.crhc.uiuc.edu> aglew@oberon.crhc.uiuc.edu (Andy Glew) writes: >>A laser might not be the best - they have a delay from zero curent while >>a population inversion is created. Normally they are used with _varying_ >>intensity rather than off/on, so there is added complexity, as well as >>the (current) need to edge couple to the fibre. LEDs are better unless >>you want coherence. Wrong! LEDs exhibit generally poor bandwidth or rise-time of the optical output. In addition, because of their broad linewidth, they are highly subject to dispersion when transmitting fast pulses over moderate distances. > > >So, what is the ballpark delay? Laser and/or LED? >Tens or hundreds of ns? > > Laser delay is a function of several factors, type of laser, modulator rise time, even output load, but 100 to 200psec is probably typical. I haven't measured any LEDS because I never considered them suitable for the reason cited above. Arnie Frisch Tektronix Laboratories
berryh@udel.edu (John Berryhill) (08/03/90)
In article <7914@tekgvs.LABS.TEK.COM> arnief@tekgvs.LABS.TEK.COM (Arnie Frisch) writes: >Wrong! LEDs exhibit generally poor bandwidth or rise-time of the >optical output. In addition, because of their broad linewidth, they >are highly subject to dispersion when transmitting fast pulses over >moderate distances. Define "moderate." I assume from the title that we're talking about distances less than a few meters. If you can detect significant phase dispersion for signals of a couple of hundred megabits per second over that distance, I tip my hat to you. It's true that LED's aren't going to run at several gigabits, but neither are the chips that you want to get the data into and out of. It is excruciatingly difficult to built a computer that uses internal data rates approaching 200 MBPS. Granted, lasers can run faster than greased owlshit, but the frequencies where designers are getting into trouble aren't all that high. I'm not going to get into the sort of argument here where we shout "Wrong!" at one another, but the problem of getting many signals around within a digital system confined to several meters is distinctly different from the problem of sending multiplexed signals through a single fiber halfway around the planet. Replacing laser diodes in a system where you've got just one fiber is no big deal. In repeaters for underwater cables, it's not uncommon to have several lasers packaged in such a way that when one fails, another can be used. But in a system where you've got lots of elements talking to one another, the low mean time to failure of lasers relative to LEDs is going to be a major headache (aside from thermal instability). You've also got to deal with rotators and polarizers in order to prevent optical feedback from the fiber from disturbing the laser. That's not a big deal for, say, a LAN or similar applications, but I'm assuming that this discussion concerns using optical interconnections internal to the box in the corner of your room with the IBM nameplate on it (or DEC or whatever). >Laser delay is a function of several factors, type of laser, modulator >rise time, even output load, but 100 to 200psec is probably typical. I >haven't measured any LEDS because I never considered them suitable for >the reason cited above. The fastest commercial LED's that I've seen have risetimes of about 2 ns. I've made surface emitters with a rise time of 560 ps and Wolfgang Harth has made ones that are < 300 ps. The question, however, is not so much how long it takes for the device to get to its equilibrium power level, but how long it takes for the optical power to cross the decision threshold of your receiver circuit and THAT figure can be mighty low. As was mentioned earlier, lasers don't work so well if you're talking about turning them on and off. That means that your receiver is going to have to be somewhat more sophisticated than, say, a simple photodiode which could be fabricated on the same chip as the digital circuit that you want to talk to. The reflexive answer than lasers are "right" and LEDs are "wrong" is reminiscent of gallium arsenide having been the "material of the future" for the last thirty years. As long as improved performance can be squeezed out of silicon, or lately out of Si-Ge alloy BJTs, then it's going to be cheaper and easier to deal with than GaAs. LEDs will always be more reliable and cheaper than lasers since they don't have to contain such a high internal optical power density or the high fields associated with maintaining population inversion. Part of the problem here, I guess, is that we've all got our pet application in mind when we argue the merits of this vs. that. Building ECL circuits is a bitch. Replacing some of the coax cable, twisted pairs, and multilevel circuit boards with optical fibers may make ECL design a somewhat more manageable bitch, but the care and feeding that lasers would require might not be worth the trouble. If we're talking about GaAs HEMT LSI chips running at a zillion bits per second with active cooling all over the place, then lasers are probably called for. I may not be right, but I don't think I'm "Wrong!" -- John Berryhill 143 King William, Newark DE 19711
ay@grad14.cs.duke.edu (Akitoshi Yoshida) (08/03/90)
In addition to John Berryhill's reference list, here are some (one intro. article, and two which theoretically talk about optical interconnects in terms energy consumption, speed, etc. (unfortunately, nothing to do with bus) Hutcheson L.D., Haugen P., Husain A., ``Optical interconnects replace hardwire,'' IEEE Spectrum, March 1987, pp. 30-35 Miller D.A.B., ``Optics for low-energy communication inside digital processors: quantum detectors, sources, and modulators as efficient impedance converters,'' Optics Letters, Vol.14, pp.146-148, 1988 Feldman M.R., Esener S.C., Guest C.C., Lee S.H., ``Comparison between optical and electical interconnects beased on power and speed considerations,'' Applied Optics, Vol.27, pp. 1742-1751, May, 1988 Department of Computer Science, Duke University, Durham, NC 27706 ARPA: ay@cs.duke.edu CSNET: ay@duke UUCP: decvax!duke!ay
davidsen@crdos1.crd.ge.COM (Wm E Davidsen Jr) (08/03/90)
In article <26481@nigel.udel.EDU> berryh@udel.edu (John Berryhill) writes: | Define "moderate." I assume from the title that we're talking about | distances less than a few meters. If you can detect significant | phase dispersion for signals of a couple of hundred megabits per second | over that distance, I tip my hat to you. It's true that LED's aren't | going to run at several gigabits, but neither are the chips that you | want to get the data into and out of. It is excruciatingly difficult | to built a computer that uses internal data rates approaching 200 MBPS. Sounds right to me. I suspect that a little led will run just fine at 100 mbit, and probably at 200 as well. I saw a prototype device sitting in a PC backplane which connected the two outer boards with an led and sensor, using a hole cut through the middle board. This bypassed the PC bus (which was only used for power and control anyway). I believe this hack was running at 50 mbit, although I certainly didn't measure it. It was carrying three million 16 bit A/D values per sec, so that sounds right. | The fastest commercial LED's that I've seen have risetimes of about | 2 ns. I've made surface emitters with a rise time of 560 ps and | Wolfgang Harth has made ones that are < 300 ps. This sounds like confirmation of what I said. We can step up a good bit from current speeds. | The reflexive answer than lasers are "right" and LEDs are "wrong" | is reminiscent of gallium arsenide having been the "material of the | future" for the last thirty years. As long as improved performance | can be squeezed out of silicon, or lately out of Si-Ge alloy BJTs, | then it's going to be cheaper and easier to deal with than GaAs. This is somewhat like magnetic media, in that something better has been just a few years away for several decades now. Magnetic bubbles and CCDs have their place, but not as replacements for mass storage. And there's always a new technology coming. I think Indium Phosphite is the one now? I don't expect to see it in my next computer, or even in my next Cray (although Cray might have GaS for some critical parts). -- bill davidsen (davidsen@crdos1.crd.GE.COM -or- uunet!crdgw1!crdos1!davidsen) "Stupidity, like virtue, is its own reward" -me