SYMPOSIUM ON "DIGITAL FILTERS"
Imperial College,
16 July 1970

Opening Address by A.H. Reeves

Mr. Chairman and gentlemen, I'm always pleased from time to time to come back to my old college. Nearly 51 years ago I entered it as a student in this same department, although not in the same building, and with nothing like the excellent facilities we have now. Nevertheless, even in my day the Imperial College Engineering Department was well known to be the best college of its kind in London.

When I had a message from Professor Cherry, asking me to give the opening talk at this symposium, quite obviously he was using the Barnes Wallace principle; he must have known, or guessed, that this is a subject I know absolutely nothing about. Now that was almost literally true a month ago. It's still nearly true - for even now I've learnt, perhaps, only 0.1% of what any speaker here ought to know before he sticks his neck out on this subject. However, I do think I've got something to say and I hope it will be of interest. It will be nothing to do with the detailed techniques of doing digital filtering, but about future computer technology in general - which could, of course, make all the difference in the world to the future of our subject today. So this is my excuse for sticking my neck out, and for being here. After only about two or three days of looking into this subject, that I know nothing about, I was really intrigued by what I saw and what I heard. First of all, it appealed to me as a matter of aesthetic elegance, to find out what we. can do by the digital computer techniques involved. You can design your filter for any pre-planned accuracy. Using 12-digit Simple Binary code, for example, this accuracy is about 1 in 4,000. But why should I waste your time? You know better than I do the various things you can do with digital filtering. One thing, though, that does appeal to me, which I'm sure will be useful sometimes, is this: by using, for example, a suitably non-linear accumulator you can build-in, accurately, non-linear characteristics with amplitude into your filter. You can of course do this by a lot of other methods as well - but for really close tolerances digital techniques will, in general be easily the cheapest.

Now as I say I've seen a very exciting future; but we've got to get things in their proper perspective. What are the main purposes of digital filtering? Obviously, they are in the same categories as filtering in general. Right, what do we need filters for? Let`s think about basic principles for a few minutes. Filtering is just one method - I repeat, one method - of sorting out, classifying, information into "boxes". Now if your information is man made as in telecom, there are three other competitive ways you can do it, but, I believe, only three. First, you can use distribution in space, which obviously is what you do when you solder separate wires to the outputs of a multi-pair cable. You also use distribution in space if you use an optically beamed method by which the channels are kept separate by differences in position or direction. Then of course you can play with polarisation, if you are using a suitable carrier frequency. You can also use distribution in voltage; some of you will remember, in the history of communication, the old quadruplex telegraph. PCM with more than two levels can again be a voltage-discrimination method. So here we have the four rival methodsof sorting our information:
1) by using the time parameter, as in FDM or TDM;
2) by spatial distribution;
3) by voltage distribution;
4) by polarisation distribution,

All this means, of course, that digital filtering is only one part of one of those four ways of doing it. Nevertheless I see very important applications, continuing indefinitely as far as I can see. Let's divide the future into two periods.

First, the next 15 years. (1970-1985)

For these next 15 years there will be, in my opinion, a really big market for FDM filters - for telecom. But how big? One obvious application is the fairly well-known 12-channel FDM system, between 60 kHz and 108 kHz. But people like the Post Office and other administrations throughout the world don't scrap things just because there is a better way. Obviously the existing analogue FDM filters on these 12-channel jobs, once they have been installed, have got to be in service for quite a long time ahead. The main use of digital filters for this application will be to meet the FDM network expansions - probably to about double the present channel-miles in 15 years time from now. After 15 years FDM production will begin to slow down, because of increasing competition from various TDM methods, particularly optical versions. When the TDM instead of FDM telecom era is fully with us, any sort of filtering either analogue or digital will be almost entirely for peripheral and subsidiary equipment, not for channel selection itself. One application will be for automatic equalisers, echo balancers, etc. Another, a very important one, will be for the automatic recognition of speech and patterns; but I'll deal with this later. In the FDM telecom, many of you know far better than I do where digital filtering techniques are likely to prove their worth. But I'll mention a few key points. First, the price of MOS chips. I would guess that in five years you will get between 700 and a thousand MOS gates on a chip, going up to ten MHz, but I won't guarantee more, for about 30 shillings. It's a remarkably cheap price per device, but digital filtering needs a remarkably large number of devices. So over the next five years, I guess, the digital FDM filters will only prove in where relatively expensive analogue filters are the alternatives (as in the 12-channel FDM system mentioned) - and even then only when there are perhaps 8 filters needed in any one spot, so that you can share the arithmetic unit and the clock. But don't worry; within these constraints there will be a large market.

Now I am going to get on to the real future. (1985 - 2000)

As to semiconductor LSI, I'm not really optimistic that within 30 years we are going to get economically beyond about 1 gigabit rates. I doubt if the yield rates will be high enough. If some new material so far unsuspected does come along, though, I could be wrong. Incidentally, if we do have a 1-gigabit rate, as far as we can foresee things now the power dissipation will be of course about 100 times what it is at 10 megabits - unless we know how we can reliably increase packing density and reduce the dissipation in each device. But I'm not sticking my neck out on this. It is likely, then, that at 1 Gb/s a thousand-gate chip for example will consume about 2.5 watts. There is no reason at all why we shouldn't, by using liquid cooling, dissipate even quite a bit more than this per chip, particularly if the chip were a bit bigger than at present.

So this is what I see on the base-band semi-conductor side. But I foresee a big and exciting future for much higher-speed computers based on optical carriers. The main basis of the optical computer of the future, as I see it, will be a new form of semi-conductor laser as we see it now, still in gallium arsenide, but in integrated-circuitry form. For laser-LSI, all we need basically is a thin film, probably of uniform pure gallium arsenide, with for example, 1000 sharply focussed spots of light impinging on it. This light, providing the. power supplies for the 1000 optically pumped lasers, and having a slightly higher optical frequency, will come from a small group of higher-power, junction-type lasers. Only where illuminated by the pumping light will the gallium arsenide film give laser oscillation. (I'm leaving out one or two important details such as the Fabry-Perot sandwich filter to increase the optical Q) Now this gives some interesting results. First, unlike junction lasers, no power needs to be dissipated at any point except where a real image impinges on the film. Second, there are no interconnecting wires, only optical beams. Third, there is also another very intriguing possibility: by the use of optical holograms, I see a way of making, of laying down, automatically, completely self-aligning optical interconnections, of up to a thousand per square centimetre. (Let me draw it.) This will be quite important, for getting rid of the fairly expensive mask making techniques now used. Suppose you have, for example, a rectangular block of transparent photosensitive material, within which you can store a solid hologram, formed by interference between two point sources of light, phase-locked together, preferably on two separate plane surfaces. In this way you can make a conjugate type of hologram, by which each point of light will give a real image at the position of the other - thus giving a 2-way optical connection between any lasing spot on one surface and a similar spot on the other. By punched tape we can put optical pumping on to any pairs of spots, in sequence, thus forming any desired interconnecting patterns. This, I-believe, will cause a big improvement in technology and also eventually considerable price reduction in computers for the highest speeds. This type of optical LSI would involve no physical connections whatsoever. The power supply is optical, on the LSI lasers, with no input or output leads in physical form at all. In fact it should be quite cheap, because you will be able to buy it in standard form, with no variations except at the hologram store, which will be a do-it-yourself hologram interconnection. I really firmly believe in this or something very much like it. I'll stick my neck out very firmly on that.

Now as to the future of digital filters in telecom after the TDM era has got into full sway, and has become optical. I foresee two main fields that are hardly exploited at all now - two fields that will expand rapidly later on. First of all speech recognition machines. I believe that they will be essential to give access for the ordinary subscriber to the information retrieval centres that I foresee in the future. I don't believe that the public will accept anything but a facility to speak their requests for information into a speech recogniser, which for cheapness will have to be automatic. Now digital filters have particularly useful applications here in the important speech analysis part of it, because of their adaptability - for example, for changing the mid-band frequencies in the filters needed to extract the speech information. Also, I believe there will be a big market for automatic typewriters. They will give a big advantage because they will provide the best of both worlds, between typing or writing it yourself when you can see what you are doing, and not having the trouble of actually learning to type fast enough - because everyone will have a solid-state TV screen on his desk on which will be displayed every word immediately after it is spoken - (or in any case, not more than a half a second later)., It will be done on a word-by-word basis, with quick means to correct a few words per page. I think that some people have over-stressed the need for context in speech recognition. It will be very important for speech translating machines. But for things like automatic typing and stylised information retrieval, I believe we can go a long way without it. I'm ready to argue with anyone about this. Another general field I can see for digital filters is pattern recognition, which is a similar problem to that of speech analysis machines. Incidentally it's rather interesting that sometimes a visual pattern can be analysed more usefully if it is turned into an audio wave and occasionally, more than occasionally, certain vowel sounds can be better recognised by turning them into spatial patterns, in polar co-ordinates. But the digital filtering techniques will be highly useful for both.

Now what I have said sounds at first sight rather a limitation to the future of digital filtering. But this is not what I mean, for these fields alone by the end of the century will have become extremely important. You will agree that the rapid and efficient spread of information is vital for the whole future of all progress. Can you think of any technological improvement more important? I can't. We can't do it properly even now, by present methods. What it's going to be like at the end of the century unless we streamline far more than we have already, is describable only by cheap limericks!