|
|||
 this log entry was all Greek to him; I guess I should have realized that many people wouldn't know what some of those terms were. So I'm here to explain one of them to you: ECL. That stands for "Emitter Coupled Logic", and it's an obsolete approach to getting high speed out of semiconductors.So let's start with some basic physics, in the form of Ohm's Laws: E=IR P=IČR "E" is voltage (in volts), "I" is current (in amperes), "R" is resistance (in ohms) and "P" is power (in watts). With respect to semiconductors, this describes (among other things) the generation of parasitic heat in the junction. A bipolar transistor is a device for switching current; you turn it on by pushing a lot of charge into the base and turn it off by pulling the charge back out again. There's also a principle which says how fast current can move into a capacitor, and I think it's I=E/C where "C" is capacitance. So the more capacitance, the slower; the higher the voltage, the faster. When a bipolar transistor is used for logic, then it's usually saturated. "On" means that the base is fully charged, which means that the resistance between the emitter and collector is negligible. By controlling R, you're controlling I, because E is fixed. Current can flow easily from emitter to collector, but because the resistance is tiny, then IČR is small, so there is no important power dissipation in the junction. When the transistor is "off", the base is fully drained. Resistance between the emitter and collector is immense, so no current flows. So again IČR is very small and no power is dissipated. So you're able to turn current flow on and off, but in both states there's no important power dissipation. Nice, huh? But when it's changing states you have a slew condition where the resistance is non-trivial and current is also flowing, and during that time IČR is non-zero and the junction heats up. However, because of the physical principles governing capacitance, that transition takes time; and it's a constant for a given transistor. E=I/C and voltage is fixed and capacitance is fixed and the amount of current needed to change state is fixed. So the amount of time needed to change states is fixed -- and pretty large (relatively speaking). To speed it up you can decrease the capacitance, which is difficult, or increase the voltage, which has its own set of problems. ECL solves the problem a different way: what ECL does is to operate the transistor in the mid region all the time. Instead of switching from all-the-way-off to all-the-way-on, it switches from partially-on to somewhat-more-on. Typically the voltage swing on the base is a fraction of a volt. As a result, IČR is always non-zero, and ECL chews power constantly and in huge amounts. No single transistor uses much, but the millions of transistors in the computer collectively will use quite a lot. (So use of substantial amounts of ECL always implies a dramatic cooling problem, and some Crays actually circulated freon from a refrigeration unit through the computer to keep it cool.) On the other hand, it means that much less charge is being put into or pulled out of the base of the transistor, so it can switch much faster. In the 1980's, ECL was the fastest logic there was. Then it was surpassed by FETs (Field Effect Transistors), which never looked back. ECL gets the maximum speed out of bipolar transistors by maximizing power consumption. Ironically, modern MOS technology gets the maximum speed out of FETs by minimizing power consumption. A FET differs from a bipolar transistor because it switches voltage instead of current. A simplistic statement, of course; both of them do both. But MOSFETs don't generally flow much current even when they are on. (On the other hand, JFETs can switch thousands of amps.) In a MOSFET the three leads are called the "source", the "drain" and the "gate". There's a channel between source and drain, and the gate sits nearby but doesn't make electrical connection. The electric field from the gate changes the characteristics of the channel and controls whether it is high or low resistance. But the capacitance of the gate is extremely small -- which means that it can be charged or uncharged very rapidly, because very little current must flow into or out of it to make it change state. No matter what they're doing, very little current flows through a MOSFET, so even when they change state they generate very little heat. IČR is always low because I is always low. The improvements in speed of MOSFETs come from decreasing their size; whenever a MOSFET shrinks, the size of the gate gets smaller and therefore it has less capacitance (which was already tiny compared to the capacitance of the base on a bipolar transistor). That means it requires less charge to change state, so it can be changed more rapidly. (It also means that they can reduce the operating voltage, which has other benefits.) And that's what the industry has been doing for the last ten years: they've been improving their technology for fabrication of MOSFETs to make them progressively smaller and smaller. That's one of the reasons why everyone is so interested in smaller and smaller IC processes. (Right now the industry is moving from 180 nanometers to 130 nanometers, with smaller processes in store.) There's a limit to this, however, because at a certain point the junction is so small that quantum tunneling means it will leak even when it's "off" -- and the more of that which happens, the less practical difference there is between "on" and "off". Anyway, that Cray which was advertised was probably built out of ECL. My 1.4 GHz Athlon Thunderbird is built from 180 nanometer MOSFETs, and I suspect it is faster. We all clear on that now? (heh) (discuss) |
|||
