- Silicon has provided the building blocks of PC chips for more than four decades, and 100GHz graphene transistors have already been shown. Ben Hardwidge investigates the future of chip materials and whether silicon's days are numbered.
- Mention silicon and most people immediately think of computer chips, or they might imagine inflated body parts that have been expanded with the silicon-based polymer silicone. Silicon is the most common metalloid in the periodic table; it accounts for more than a quarter of the elements in the Earth's crust and we've depended on it to propagate the computing revolution over the past few decades.
- Computer chips don't have to be made from silicon though. The nature of binary is that it only requires a device that can be switched on or off. You could perform the same task using a binary switch made from all sorts of other materials: In fact early binary computers such as the 1944 EDVAC used thousands of vacuum tubes and mercury delay line memory; The main reason we use silicon is because it still functions as a semiconductor at high temperatures, making it ideal for packing millions of fast-switching transistors into a densely packed space.
- Intel has been making processors from silicon since it first introduced the 4-bit 4004 in 1971, but does a potentially better material exist? In February 2010, scientists at IBM announced that they had successfully demonstrated a 100GHz transistor made from a material called graphene. The Internet gossip circuit was soon awash with predictions of terahertz graphene processors and silicon's funeral march; this is hardly surprising given that IBM's own press statement claimed that the transistors could be featured in 'zippy computer processors' in a decade's time.
- We'll come straight to the point before we start singing the silicon swansong though – graphene won't fully replace silicon as the main material used for CPUs - at least not yet.
- IBM scientists from the company's TJ Watson Research Center have demonstrated the operation of graphene field-effect transistors at 100GHz frequencies.
- Yu-Ming Lin, a researcher at Nanometer Scale Science and Technology at IBM explained, 'There's an important distinction between the graphene transistors that we demonstrated, and the transistors used in a CPU'.
- 'Unlike silicon,' says Lin, grapheme doesn't have an energy gap, and can't be "switched off", resulting in a small on/off ratio.' He also said that the 100GHz frequency reported was the cut-off point (the highest frequency at which the transistor operates) rather than the target frequency of a transistor, which would be around a third of this. IBM says that this is more than double the cut-off frequency of current top end silicon transistors, which is currently around 40GHz.
- Graphene will still be a very important material in the future of integrated circuits, though, and this includes CPUs, even if it doesn't fully replace silicon.
- Lin points out that the on/off ratio isn't an issue in RF circuitry, for example, and it has lots of properties that potentially make it preferable to silicon, depending on the circumstances.
- Looking like the cross-section of a Crunchie bar, graphene groups carbon atoms and their bonds into a lattice of tiny hexagons. And when we say tiny, we mean as small as possible without splitting the atom. The length of one carbon bond is just 0.142nm, and the material is basically two-dimensional at the atomic level, meaning that a sheet of graphene has the thickness of a single carbon atom. As a point of reference, a 1 mm-high block of graphene would contain more than three million sheets.
- Does this mean that we could create transistors at the atomic level? It does indeed. In fact, Dr Kostya Novoselov and Professor Andre Geim from The School of Physics and Astronomy at The University of Manchester did exactly that in 2008, when they created the world's smallest transistor from graphene; it had a width of ten atoms and the thickness of just a single atom.
- 'In principle, there's no limit to the size of a graphene transistor,' says Lin. 'Compared to silicon, graphene is more robust in terms of device scaling, because it has the thickness of a single layer of atoms, while it's known that the quality of silicon will suffer significantly once it's thinned down.'
- Size and scalability isn't the only benefit of graphene either. 'Graphene possesses a much higher thermal conductivity than silicon,’ explains Lin. 'Therefore, it can transport and dissipate heat more efficiently than silicon. More importantly, unlike silicon, the performance of graphene devices is less temperature-sensitive, indicating that they can operate within a larger temperature range. These factors would also mitigate some of the power requirements we currently face with silicon.'
- We therefore have a material that's scalable to the atomic level and less susceptible to heat than silicon. What’s more, as IBM's 100GHz cut-off frequency demonstrated, graphene transistors can also be clocked far higher than comparable transistors made from silicon. Not only that, but in September 2010, researchers at the University of California, Los Angeles (UCLA) reported that they had achieved a phenomenal graphene transistor cut-off frequency of 300GHz.
- So why does graphene offer such massively higher clock speeds? 'Graphene transistors can achieve a higher clock speed (or frequency) than those made of silicon with the same length because the electrons in graphene can move at a higher speed than those in silicon,' explains Lin. 'The speed of a transistor is basically determined by the velocity of electrons, and is inversely proportional to the length of the gate.’
- If grapheme were capable of the same on/off ratio as silicon, it would probably entirely usurp silicon in the world of computer chips once it was past the research phase. As it is, it's likely to replace silicon in some areas but not in others – at least not in its current research state.
- 'Graphene as it currently is won't replace the role of silicon in the digital computing regime, confirms Lin. However, it may complement silicon in the form of a hybrid circuit to enrich the functionality of computer chips.'
- Slicon is one of our planet's most abundant elements, but creating a compound of two rarer elements, and then fabricating processor parts from them, automatically pushes up the cost of manufacturing.
- Rather than silicon being completely replaced, we're more likely to continue to integrate new materials into chips, and silicon itself, in order to improve them. Silicon may not be perfect, but it was chosen as the default chip-building material for a reason. It's the most appropriate material for the job, and until recently, only a few tweaks have been required to keep it in shape.
- Intel’s director of components research, Mike Mayberry, told us that, 'silicon's properties make it a nearly ideal material'. He also pointed out that 'the industry has so much experience with it that there are no plans to move away from silicon as the substrate for chips'. The substrate is the material on which the logic gate of a transistor sits, and is separated from the gate via an insulating material.
- However, while silicon is likely to remain the standard material for the substrate of chips, there's still plenty of room for experimentation with other materials in different parts of the chip, including the transistors. 'We're actively investigating the addition of new materials to silicon to make better transistors,' says Mayberry, offering the examples of high-k dielectrics such as Hafnium oxide (HfO2) and III/V semiconductors such as gallium arsenide (GaAs).
- Let’s start with the latter. Like graphene, gallium arsenide has a major advantage over silicon in the form of its higher electron velocity, enabling transistors to run at higher operating frequencies. In fact, back in 1969, the original patent application for a GaAs semiconductor device from a pair of Japanese inventors Takeshi Saito and Fumio Hasegawa, stated that GaAs diodes could have a high frequency of ‘over 300GHz'.
- It’s also less sensitive to heat, which makes it preferable to silicon for transistors. Unlike silicon, GaAs is a compound material made from two elements — gallium and arsenic. It's called a III/V material due to the parts of the periodic table from which its fundamental elements are grouped - III materials are in the boron group and V materials are in the nitrogen group.
- Gallium arsenide has already been used in integrated circuits for decades, and the Cray Group even used it to build a supercomputer processor in the 1990s. However, it's only recently that Intel has started to talk seriously about it being used inside its CPUs.
- Part of the problem with introducing new materials for chips is the cost. You don't generally find blocks of silicon occurring in nature, but it's one of our planet's most abundant elements — you can easily find it in sand, for example. Conversely, creating a compound of two rare elements and then fabricating processor parts from them automatically increases the cost of manufacturing.
- For this reason, silicon will be used as the default material for substrates in computing chips for some time, but Intel is keen to introduce new materials for various other parts of the CPU. Last year, Intel reported high-k dielectric results from InGaAs (indium gallium arsenide), and this year it also reported the results using germanium (Ge — a group IV material like silicon). According to Mayberry, 'both results show a possible path forward'.
- 'The silicon itself isn't limiting,' explains Mayberry, 'but there are challenges to be overcome with shrinking feature sizes. Examples are lithography, interconnect speed, memory cell size, transistor performance and energy efficiency.'
- ’While silicon remains the substrate for manufacturing,' says Mayberry, 'various components are being replaced. For instance, aluminium interconnects have been replaced by copper/silicon dioxide dielectrics, while polysilicon gates have been replaced by high k/metal gate materials at Intel.'
- Meanwhile, hafnium oxide is already being used as a high-k dielectric in Intel's current line-up of CPUs. Basically, instead of a transistor featuring a silicon logic gate that's separated from the silicon substrate via a silicon dioxide insulator, a high-k/metal gate transistor features a metal gate and a high-k dielectric insulator.
- The main advantage of using a high-k dielectric material instead of silicon is the strength and thickness of the material. Silicon dioxide transistors have major problems with leaking current as they're made progressively thinner, but thicker; high-k materials don’t leak as much, making them an ideal material as transistors become smaller.
- This is a prime area in which silicon is likely to be completely replaced. 'This thicker class of materials, known as "high-k", will replace today's silicon dioxide technology,' says Mayberry, adding that this will ‘then provide extendibility over several generations’.
- IBM's new CMOS Integrated silicon Nanophotonics chip technology integrates electrical and optical devices on the same piece of silicon, enabling computer chips to communicate using pulses of light instead of electrical signals.
- Aside from the transistors themselves, the interconnects between the components of a processor are also rapidly changing in terms of the materials they use.
- Copper is currently used to create interconnects inside CPUs and there has been research conducted into using nanowires as interconnects (more on this later). However, one potential future development here is the use of optical connections.
- We already use fibre optics for everything from audio connections to broadband and TV, but optics are also making their way into computers as high speed connections. Intel's Light Peak technology has already been in the spotlight as a potential replacement for USB and FireWire.
- Offering a bandwidth of 10 Gb/sec in both directions, which Intel says will rise to I00 Gb/sec within ten years of launch, Light Peak could potentially revolutionise external devices.
- According to Intel, the technology could enable you to transfer an entire Blu-ray movie in less than 30 seconds.
- Intel's ambitions for optics don’t stop there though. We caught up with Intel's strategy and business initiative director for its Photonics Technology Lab, Jeff Demain, who predicts that optics could become much smaller and result in super-high-speed interconnects not only inside a PC, but also inside a processor.
- ’The concept of optics within a microprocessor for high speed interconnect certainly has merit,' Demain told us. 'An industry path one could envision towards this goal would start with optics outside the system (today's active cable), then optics co-packaged within the system, and finally optics within the microprocessor.'
- We could wait some time for this, but it isn't outside the realms of possibility. 'The combination of technology advancement and business factors will set the timing,' says Demain. 'This will make full integration a viable option in the long rather than short term.'
- Research into optical processor connections, dubbed nanophotonics, is already in full flow. In December 2010, IBM unveiled a chip that featured both electrical and optical devices on the same piece of silicon.
- 'With optical communications embedded into the processor chips,' said Dr. T C Chen, vice president of science and technology at IBM Research, the prospect of building power-efficient computer systems with performance at the exaflop level is one step closer to reality.'
- Not only that, but IBM also says that the process involves 'adding just a few more processing modules to a standard CMOS fabrication flow', resulting in ‘a variety of silicon nanophotonics components'. The keyword here, of course, is silicon.
- As such, integrating optical interconnects into a processor could be much cheaper than introducing new compound materials such as gallium arsenide to the mix. ‘Single–chip optical communications transceivers can now be made in a standard CMOS foundry,' explains IBM, 'rather than assembled from multiple parts made with expensive compound semiconductor technology.'
- Optical connections aren't the only future possibility when it comes to internal interconnects in a processor. Another potential development involves nanowires, a term that encompasses a variety of definitions, depending on who you ask.
- Some people claim that a nanowire has a width of just a single nanometre or less, but it’s generally accepted that the term also be used to describe wires with a width measured in tens of nanometres.
- Nanowires can also be manufactured from a variety of materials, including metals, insulators and semiconductors. In terms of making circuits smaller, nonowires clearly have many benefits, but they also present a challenge, as this is the scale at which quantum mechanical effects are observed. Investigating quantum mechanics1 in depth is beyond the scope of this feature, but this is basically the point at which it becomes difficult to make accurate predictions about how a material will behave.
- Nanowires are still very much in the embryonic2 stages of research, but this isn’t to say that they can't play a part in future CPUs. We interviewed Intel's technology analyst, Rob Willoner, about their use in processors. He reckons that ‘nanowires have a potential application as either interconnects or transistor channels’, although he added that ‘they're very much in the research phase'.
- According to Willoner, part of the problem with implementing nanowires is the sheer complexity of miniature integrated circuits. He offers the example of carbon nanotubes (a type of nanowire). ‘These are grown from a seed catalyst, and have very daunting placement challenges when you consider trying to hook up a billion devices.'
- lt’s also worthwhile talking about another exciting development in computing – spintronics. This process uses the spin of electron to represent binary ones and zeroes as in terms of storage space.
- The 'spin' of electrons was photographed last year by researchers at the University of Hamburg. The image showed that individual cobalt atoms appeared as a single protrusion if the spin direction was upward, and as double protrusions with equal heights when the spin direction was downward.
- The actual 'spin' is a quantum-mechanical property of electrons, and in spintronics, it can be effectively used to send a signal to a transistor without any charge. Normally, a transistor requires a current between the source and the drain (on either side of the gate) in order to operate. However, using spintronics, a transistor can instead obtain its data directly from the spin of an electron.
- The theory of spintronics is already used in some computing devices, such as the GMR heads in hard drives, but we're still a long way from seeing it used in processors. In fact, at the 2008 Intel Developer Forum in Taipei, Intel stated that it would be at least another ten years before spintronics made its way into computer chips, which means we still have at least seven years to wait.
- Even so, research into spintronics is forging ahead at an amazing rate, and last year the 'spin' phenomenon was even photographed by a team of physicists at the University of Hamburg. By potentially limiting the need for charge, spintronics could result in chips producing much less heat, which can only be helpful given that transistors are being packed more densely into chip packages.
- We clearly have some exciting times ahead of us when it comes to new chip materials. After spending a few decades simply packing in smaller and smaller transistors, the weaknesses of silicon are starting to show and resolving the problem requires some really interesting innovations.
- While silicon is likely to remain an important material in CPU production for the foreseeable future, experiments with graphene have demonstrated the potential for new transistor materials, and we're likely to see graphene replacing silicon in other types of integrated circuit in the future.
- In fact, further development of graphene could result in completely new devices that can take advantage of its ability to run at higher frequencies without needing the larger on/off ratio provided by silicon.
- 'Graphene is still very much in the research phase,' explains Intel's Mike Mayberry, 'but researchers are predicting a number of interesting properties for it and experimentalists are trying to confirm them.' According to Mayberry, these 'may translate into new applications that don't look like conventional devices, with further research'.
- We probably won't see the 300GHz3 graphene CPU that many people predicted anytime soon but there are some amazing developments occurring at the atomic level.
- The fact that we can produce a transistor with the thickness of a single atom is a great achievement, as are the concepts of spintronics and nanowires.
- Silicon is here to stay for the moment, but more experimentation is required to keep it going as transistors become thinner and leak more current. We're likely to see more semiconductor materials introduced to the mix over time, and we could also see the integration of other ideas such as optical interconnects and carbon nanotubes in future processors.
- With so many innovations focused on reducing the power consumption and heat output of transistors, while also potentially increasing clock frequency, tomorrow's chips are not only going to be faster, but also cooler and smaller.
Footnote 3: The article had 30GHz.
Text Colour Conventions (see disclaimer)
- Blue: Text by me; © Theo Todman, 2020
- Mauve: Text by correspondent(s) or other author(s); © the author(s)