henry@utzoo.UUCP (Henry Spencer) (06/14/83)
A friend of mine objected to one statement I made in my recent note about why Seymour Cray doesn't use nontrivial ICs. His objection, and my comments in reply, may be of interest as an expansion on my earlier note: ----- What do you mean "line termination within IC's". There's no such thing. When you're sending a signal 3 millimetres across a chip, you don't need to terminate it. Signals between chips still need to be terminated at the chip, of course. ----- What do I mean "line termination within IC's"? I haven't the foggiest idea. It sounds a bit odd to me too. I am more-or-less quoting from an article in IEEE Computer, and it's possible that the writer got his facts wrong. But I'm not Seymour Cray, and I do know he has problems far beyond those that mere mortals encounter. He needs proper line characteristics for reasons above and beyond avoiding reflections -- the power demand from his ICs must be as close to pure DC as possible, because he'll get standing waves in his ground planes if the power demand is switching in time with the clock signals. This is just as important a reason for using differential pair everywhere, and may account for his objections to MSI (it's been a while since I read the article in question). Henry Spencer U of Toronto
newman@utcsrgv.UUCP (Ken Newman) (06/14/83)
I read that due to complete symmetry of some kind in the Cray-1 circuit modules (there are only 4 different types as I recall) the load on the Cray power supply is purely resistive!!! The crazy thing still consumes 115 kW though (good old ecl). I can find the article if anyone cares.
smh@mit-eddi.UUCP (Steven M. Haflich) (06/15/83)
Last week I was visited by an old friend who now works in Gallium Arsenide (GAs) logic research for a large west coast company. Unless I have my figures entirely scrambled, he told me that work is proceding on ECL compatible gate array logic operating at gate delays approaching 50 picoseconds. (Well, he did admit that the gates were only probabilistically digital at these speeds, but the bugs were being worked out.) In case you have forgotten, 50 picoseconds fits into a microsecond about 20,000 times! Anyone who has dabbled in microwaves will realize that such frequencies give new meanings to terms like "parasitic capacitance" and "transmission line effects." When one builds a radio transmitter operating at a mere couple hundred MHz, one must give careful consideration to the shape and path of simple conductors connecting circuit elements; at such frequencies, there ARE no simple conductors! Now imagine that the conductors have to be made of/in epitaxial GAs, and must operate at two orders of magnitude higher frequency. It is no coincidence that my friend got into the GAs business directly from microwave work. Don't bother asking me for more info -- I have pretty much exhausted my understanding of the field. However, it appears that swallowing IC's will soon be even a worse idea than it is now. Steve Haflich, genrad!mit-eddie!smh
padpowell@wateng.UUCP (PAD Powell[Admin]) (06/15/83)
The whole area of high performance digital logic borders on black art, with a healthy dab of applied mathematics, mechanical engineering, and love of experiment. For instance, if you send a signal down a line, it will keep on going until it comes to a discontinuity. Whereupon really ugly things will happen. The usual analogy is that of a wave in a bathtub. Start the wave at one end, and watch what happens as it hits the other: reflections, etc. To keep this crud to a minimum in electronic circuits, it is neccessary to terminate a line with the same impedance. Lines are usually designed to be purely resistive, so terminating them with a resistor is called for. Now think about all those little chips, about 3mm across, needing all those huge resistors... So the cunning chip designers put the resistors on the chip, making the chip slightly larger, and increasing its power consumption. Unfortunately, this is only the first step. Due to other problems, it is usually not possible to run a single line for a signal. It becomes neccessary, for distances over about 4 inches, to use two lines, and drive them differentially. The design of the hardware for this monster has to be seen to be believed. Strangely enough, I have the feeling that it is one of the last areas to be impacted by the CAD market... After all, who would invest the hundreds of man years neccessary to design a system which would automatically lay out a board? Answer: IBM, Japan Inc., ITT, and other people... Hot topic for people interested in wierd and wonderful CAD applications. Patrick Powell, U. Waterloo
anton@hcr.UUCP (Anton Aylward) (06/15/83)
Henry (utzoo!henry) although quoting, is perfecly correct. The "whadya mean termination its only 3 mils..." class of argument doesn't cut any ice. What is critical is the rise time of the waveform. Suppose you have a 100 Megahertz square wave clock: Unless you have a bandwidth above 20 times that the 'fiiltering' is going to stop it looking like a square wave. OK HAPPY. Kids stuff from first year. But now you realise we are working with microwave freequencies. I know from my days working in Silicon Valley for a "semiconductor firm who shall be nameless" on military high speed LSI and ECL at these speeds, the chip designers worry about transmission line effects. Hey, these guys worry about the transmission line effects of power surges on the power input line. If you look, you will find many chips have power feed on pins at the middle of the chip rather than at the end. Now start facing some reality: the signal doesnt travel along the conductor, it travels in the gap between the conductor and the groundplane. Where is the groudplane on most chips ? Somewhere a few tens of mills over there. The length of the line is irrelivant under these conditions. What zonks it all out of exitence is the dialectric (say he avoiding be Hegelian). This is fairly grooty for microwave freqeuncies and so produces significant wave dispersion. This means wave form dispersion, and differential speed of transmission with freqency. Plop a signal along something short that looks like this and what hits the end, never mind what bounces back out at you, doesn't look like the square wave you shoved in. ------------------------------------------------------------ WHY WHY WHY is it that software bods and even the bulk of hardware bods just don't realise you cannot ignore microwave effects, termination and the rest when designing switching circuits ? Even at the 5 MHz the noddy micros run at these effects are noticeable. At the 10 to 15 MHz of the new 16 bitters they are starting to become significant. FACE REALITY guys, or you'll go the way of FORTRAN. /anton aylward
dmmartindale@watcgl.UUCP (Dave Martindale) (06/15/83)
I was under the impression that Cray used only very simple IC's where all of the emitter outputs went off-chip and thus could be tied to resistors to give the constant power consumption regardless of output state that he desired. More complex MSI chips would have internal gates which fed only other internal gates and thus the loading of these internal gates could not be controlled.