If I mention 'spin' in this column nowadays you'll probably expect me to be talking about politicians getting macro-economic with the truth, so it's refreshing to be reminded that the word had a more old-fashioned meaning, namely a property of things that go round-and-round such as tops, wheels or electrons. What reminded me was Scientific American's Special Issue called 'Beyond Einstein' (Sept 2004), which anticipates next year's centenary of the Special Theory of Relativity and contains an article on its practical, er, spinoffs, one of which is the newish field of solid-state physics called 'spintronics'.
Regular readers may recall that last year I wrote a feature in this magazine about the Physics of Computing, in which I concluded three main points: that we owe the PC revolution to the benign scaling properties of CMOS semiconductor fabrication technology; that this virtuous circle may soon be broken by problems of heat dissipation; and that CMOS's hotly-tipped successor, the quantum computer, as yet lacks any similarly-credible fabrication technology. The good news is that spintronics, rather than quantum computers, may provide the solution we need. The bad news is that, though spintronic devices will be faster than CMOS for a tiny fraction of the power, they won't exhibit the semi-miraculous superpositions (and hence massive parallelism) of quantum devices. Better news is that they'll exhibit other neat tricks, such as multi-valued logic and the ability to morph one circuit into another under software control.
First back to basics. Our present computers, cellphones, GPS, iPods and all the rest, are 'electronic' devices that work by squirting large numbers of electrons from one place to another according to the dictates of Ohm's Law: it's a bit like plumbing, with electrons as the water and voltage as the head of pressure. Electrons are shunted around in bulk so their quantum properties don't figure at all. Spintronic devices on the other hand work by measuring the spin properties of small numbers of electrons. When an electron is exposed to a magnetic field its axis of spin 'precesses' about the field direction - much the way a spinning top wobbles as it slows down - and this effect can be measured by various electronic or optical means, analogous to those used in medical Nuclear Magnetic Resonance Scanning.
So Far Not So New, because spintronic devices that employ ferromagnetic materials to provide the magnetic field have been known for some while. IBM's successor to the Giant Magneto-Restrictive (GMR) hard disk head is based on a spintronic device called the Magnetic Tunnelling Junction (MTJ), while IBM, Motorola and others have demonstrated Magnetic RAM (MRAM) chips based on the MTJ or similar devices called 'spin valves'. MRAM is destined to become very important indeed because it combines the best features of Static RAM (it's non-volatile and low-power) with those of DRAM (it's fast and cheap to make) and all-being-well it ought eventually to replace both DRAM and Flash RAM, giving us 'instant-on' PCs and MP3 players with a battery-life measured in months. However these technologies have a serious flaw as candidates for making microprocessors, namely that they require the deposition of layers of ferromagnetic materials like barium titanate, and that doesn't mix well (or indeed at all) with CMOS fabrication.
Here's where Einstein comes into the picture: Special Relativity predicts that when electrons move through a static *electric* field they experience part of the force as if it were a *magnetic* field. In other words you can impose a magnetic field without having to deposit permanent magnets on a chip. And of course static electric fields are precisely what CMOS chips deliver, which is why we call them Field-Effect Transistors. The effect is very small unless the electrons are moving very fast, but then electrons in modern CMOS chips do move fast, at significant fractions of light speed. David Awschalom and his team at UCal Santa Barbara produce a strained crystal lattice by layering two slightly different semiconductors, generating an internal electric field that causes electrons' spin axis to precess and to stay put once they leave the field. The field's peaks and troughs herd electrons in a desired direction, and varying the voltage (to change the electron's speed) or the geometry of the strained-lattice 'wires' gives enough spin control to encode 'spin-bits'. For now they use lasers to read the spins, but the next step is to create a whole 'spin transistor' that reads and writes bits within the same chip, and can be made by a conventional fabrication process.
There are several exciting things about spin bits. When the field is removed they stay put, so that no power at all is consumed between operations. They can be switched in picoseconds. They're coherent, so you could encode data using their relative phase angle rather than just up/down (rather the way ADSL does) permitting more values than just 0 or 1 to be stored in each 'bit'. And a team in Berlin has just shown that spin logic gates can be built using fewer transistors than CMOS, and can change their function from AND to OR to NAND (or whatever) under software control. We're looking at the possibility of processor chips faster than today's and packing 10,000 times more data for only 15% of the power consumption. That's a huge step closer to the efficiency of the human brain, which crams in around 1,000,000,000,000,000 processing elements while still dissipating less heat than a Pentium 4...
No comments:
Post a Comment