Half metals and other new materials hold the key to future semiconductor scaling.
For close to 40 years, the semiconductor industry has disliked adding new materials to chips. They often cause problems that take several process generations to resolve. A silicon dioxide gate on top of each transistor in a design would today be considered 'tradtional', but the technology took years to perfect. So, there has been little enthusiasm for replacing it.
Now, though, it is giving way to more complex metal-oxide gates in applications beyond high-performance microprocessors. Further, having introduced rare-earth metals such as hafnium to improve end-performance under the CMOS manufacturing process, chipmakers are exhibiting a new confidence about their materials choices.
As you may have guessed, the driver for this reinvention is necessity. Having almost exhausted the potential of silicon, this decade could demand the adoption of new materials, according to scientists and engineers such as Intel's senior components researcher George Bourianoff. 'Scaling will continue. But performance improvements will come from new materials and new structures,' he says.
Near-term changes are likely to be evolutionary, tracing the footsteps of techniques that have already been used to gain improved performance from strained silicon. But, as 2020 approaches, there could be a radical shift away from silicon in a quest for materials that can support electrons that behave as though they are massless - hugely improving conductivity when circuits are active.
The trend in new materials started with the reintroduction of germanium over the past decade. An element that was used by early researchers - including Jack Kilby at Texas Instruments - germanium interferes less with the movement of electrons moving around the lattice in the material's conduction bands. This means less resistance and less power loss. Unfortunately, the element is not easy to work with, and it has taken three decades to find a stable material to act as a gate on top of a germanium transistor.
More recently, by mixing germanium with silicon, companies found they could achieve a speed boost from the way that the larger element stretches a conventional silicon crystal lattice. Dimitri Antoniadis of the Materials Structures and Devices Focus Centre at Massachusetts Institute of Technology (MIT) points out that this type of material change was necessary because, once past the 130nm process node, simply reducing the size of a transistor did not allow it to go faster - resistance to electron flow became a big problem. So, process engineers have had to add more dopants and crystal-stretching features.
'Transistor characteristics have improved but only through the addition of new technologies such as strained silicon,' says Antoniadis. 'There has been an increase in electron velocity that has allowed scaling to continue that, left to its own devices, silicon could not have achieved.'
The problem now is that even with the large amounts of added germanium and other structures to strain the silicon lattice, there is not a great deal of room for improvement.
III/Vs and carbon
One option is to use silicon merely as a strong and stable substrate for wafers and move elsewhere in the device to materials that have always been faster. A set of candidates here is based on elements in Groups III and V of the Periodic Table (silicon and germanium are in Group IV). Communications engineers are already using III/Vs such as gallium arsenide for high-speed transceiver chips because of their high speed and low power-consumption. However, III/Vs have not so far been used more widely because they are very fragile.
Intel, defence technology group QinetiQ and academics such as Antoniadis are working on ways to reliably put these materials on top of a silicon wafer. One problem they face is that the lattice mismatch is so great that the III/V layers simply crack. Test devices from Intel and MIT are composed of many graded layers that attempt to minimise the stress as the lattice shifts from silicon to III/V. The gains are apparently dramatic: the same speed for ten times less power in some cases. However, unless engineers find a way to put a stable III/V layer directly on top of silicon, such devices will be expensive to make and therefore very limited in their viablity.
The next step may be to move to carbon as a conductor, using such forms of the element as nanotubes and graphene. Researchers at the University of California at Los Angeles (UCLA) recently demonstrated a transistor made from graphene that could switch at 300GHz. The team, led by Professor Xiangfeng Duan, put a nanowire on top of a graphene layer in an attempt to work around one major problem with all carbon-channel devices: getting them in the right place.
'This new strategy overcomes two limitations previously encountered in graphene transistors,' Duan claims. 'First, it doesn't produce any appreciable defects in the graphene during fabrication, so the high carrier mobility is retained. Second, by using a self-aligned approach with a nanowire as the gate, the group was able to overcome alignment difficulties previously encountered and fabricate very short-channel devices with unprecedented performance.'
However, the devices made so far using graphene or nanotubes have been isolated transistors where alignment and position are not critical. MIT's Antoniadis points out that there are no known solutions for integrating graphene or nanotubes.
Even with the higher mobility of III/V transistors and potentially graphene, there are problems. Antoniadis points to resistance at the contacts as a big hurdle. 'And that can easily take away 20 per cent of the device performance,' he says.
Interfaces in silicon have always been trouble. One huge problem that held up the development of the high-k metal gate structures now routinely used to replace silicon dioxide was getting manufacturable and compatible materials to act as low-resistance contacts. Bad interfaces also result in vacancies and impurities that trap charge carriers and disrupt the operation of the device.
A better understanding of materials may make the interfaces key elements in the next generation of high-speed, low-power devices. Engineers may draw on concepts from the further reaches of quantum-mechanical theory to create new types of transistor, even dynamically tunable ones. These devices might invoke quasiparticles still more exotic than the holes that feature heavily in semiconductor physics. Rather than simply exhibiting the absence of a charged particle, these quasiparticles are the result of interactions between basic particles that, under the right conditions, move predictably and can be controlled.
Such materials might be said to have multiple personalities. Their behaviour depends on shape and surface properties, so they can be, at once, metallic, insulators or semiconductors. Graphene provides the inspiration for these materials thanks to its unusual electron orbitals that result from its lattice of repeating benzene rings.
The curious nature of the benzene ring intrigued scientists for more than a century. In principle, the ring should consist of alternating single and double carbon-carbon bonds. The single bonds are formed from electron orbitals that in effect point towards each other when they overlap. The second bond is the result of orbitals that are parallel to each other but close enough to overlap to form a new bond. This stronger, double bond means the two carbon atoms are closer to each other than they would be in a single bond. Centuries ago, scientists realised that benzene is a regular hexagon, not the distorted one you would expect if the molecule had alternating single and double bonds. Something else was clearly going on.
Rather than being isolated to individual bonds, the electrons become delocalised over the entire layer. So, as well as the normal direct, single bond between carbon, there is a layer of delocalised electrons above and below the carbon atoms made up of interlocking rings. In this layer, the electrons can move as though they are in a metal.
Where chemists talk of delocalised electrons, physicists use the idea of the Dirac cone: a zone where electrons have no effective mass. Stephen Jenkins of the Surface Science Group at the University of Cambridge says: 'Graphene became really interesting in 2003 and 2004 when people found they could make it readily. It was the only material at the time where people had seen Dirac cones. But now there are topological insulators where these cones are seen.'
In the topological insulator, the core is an insulator but Dirac cones form at the surface and provide a highly conductive layer.
Jenkins has uncovered a third type of material with a split personality that may offer a wider range of ways of controlling electrons. Although it is not yet clear exactly why, the alloy of nickel, manganese and antimony (NiMnSb) has a lot in common with graphene. 'It's the symmetry of graphene that makes things happen, and that's also the case with this alloy,' Jenkins says.
NiMnSb is an example of what is known as a half-metal - only electrons in the right spin state can move easily through the material. For electrons in the opposite state, the material is an insulator. It is not yet clear whether the half-metal has a split personality like a topological insulator but Jenkins thinks that is likely.
'Potentially, we've got a region that is at the surface of this material that, like graphene, has exciting properties. And it is sat on top of a bulk material that has the exciting properties of a half-metal,' he says. 'It's something that a device specialist might think of interesting things to do with.'
Interest in half-metals goes back to 1983 when the concept was first proposed by researchers led by Professor Robert de Groot of the University of Groningen.
Because electron flow depends on spin state, it opens up the possibility of using the core of a material like NiMnSb as an electronic device based on spin states rather than changes in voltage or current. 'You can keep the current flowing and flip the electrons up and down: this happens an order of magnitude faster than where you are switching current on and off,' says Jenkins.
Spin-dependent devices already exist, such as magnetoresistive memories. But Jenkins points out that they often need to be composed of complex sandwiches of materials laid down atomic layer by atomic layer. 'The interest in half metals is that maybe you can jump to this simpler type of material where only one type of current flows,' he says. 'With other materials you also get surface-localised states where you would hope to only get spin-up electrons but you get both types. Those minority states can be a problem. In other materials they are a problem but with this you have a surface that has Dirac cones, which offers other possibilities.'
There are subtle differences between the highly conductive parts of graphene and NiMnSb. 'The Dirac cones [in NiMnSb] are in less symmetrical positions. They are both tilted and elliptical, where in graphene they are upright and circular,' says Jenkins.
He is conducting research into sandwich structures for the half-metal alloy to explore what happens at interfaces between layers. 'It's getting closer to a device-type architecture, where you could apply electric fields across the spacer and see electrons jump from one side to another. This is where the details of Dirac cones become important.
'There are all sorts of materials science problems to be solved. What I am doing is pointing up interesting properties where device scientists might get involved,' says Jenkins.