Original URL: https://www.theregister.co.uk/2012/05/08/big_data_chip/
Boffins baking big-data single chip architecture
Graphene, electrons and the end of 'conventional silicon electronics'
Some use software – caching, in-memory transactions or BigTable-style algorithms to cluster and control groups of servers. For others, the answer lies in the hardware: packing more cores into chips or making the transistors faster. Both schools are looking for ways to make applications, computers and servers capable of processing big volumes of data without cramping up.
US university researchers reckon these techniques have had their day, however, and have turned to nano electronics for the answer.
A team led by scientists from the University of California have returned to electrical engineering fundamentals with work that will produce a chip combining memory and logic. Only this will overcome what they call the “ultimate limits” of conventional silicon electronics the others are trying to circumvent. Today, you have memory in RAM and logic in the CPU connected by a bus that causes a bottleneck as instructions are relayed.
Their goal is to construct a functioning magnetologic gate that will be the building block of the new chip architecture and become the main element of a circuit.
The idea is to emulate the transistor, which paved the way for the integrated circuit and microprocessor – foundation stones of today's computers. The magnetologic gate will be used on circuits in intensive applications like search, data compression and image recognition.
The physics and astronomy – yes, astronomy – professor leading the project is Roland Kawakami, who told The Reg in an interview he reckons the magnetologic gate is just three years away from becoming reality.
“It’s completely rethinking how you do computing,” Kawakami told us. “We are in a crawling phase [now]. It’s similar to back 50 years ago, when they invented the transistor and they needed that one device. That’s the same for us.”
He’s been working on the ideas around this since 2010. His research spun out of work for DARPA, but about six months ago the project went live and landed a helpful $1.85m grant from the National Science Foundation and the Nanoelectronics Research Initiative whose members include IBM, Intel, AMD and Texas Instruments. The cash will fund 14 researchers and experts on five campuses. Kawakami is at UC Riverside and he will be joined by researchers at the University of California Irvine and the University of California San Diego as well as two universities in New York State whose expertise spans magnetoresistive memory, theoretical physics, circuit design, and constructing integrated circuits.
At the heart of the project is graphene, the wonder material that became popular in 2010 following the Physics Nobel prize-winning work of the scientists who discovered it: the University of Manchester’s Andre Geim and Kostya Novoselov. Graphene has opened up new doors because it's super lightweight, thin and capable of conducting electricity.
The magnetologic gate uses a set of magnetic electrodes that are connected via graphene. Binary data is stored in the electrodes' magnetic state (ie: north and south produce zeros and ones respectively), while the logic is determined from the electron's spin state in the graphene.
Electrons and graphene were made for each other, as the former moves through the latter faster than it does through silicon – at 1/300th the speed of light, about 10 times as fast as electrons in conventional silicon devices – while its polarisation rate increases 30 per cent when compared to regular semiconductors.
Graphene saves the day
The work builds on earlier breakthroughs: the idea of spin-based computing using a magnetologic gate that was devised at UC San Diego in 2007 and on tunneling spin injection and spin transport in graphene from a group led by Kawakami in 2010.
In 2010, Kawakami’s team realised that graphene was good for helping create electron spin with the help of a magnet working at room temperature. This was a vital breakthrough because it put the economics of constructing an affordable device and physics of making it work into the realms of the possible.
Kawakami said: "After our breakthrough on performance in 2010, we ran numbers we got on graphene and found it could run very fast, at the gigahertz level, and the amount of power used would be very low. This is a potentially viable technology... by integrating the memory with the logic especially for any applications that use lots and lots of data, like search, image recognition and compression."
But the hard work of bringing all that work and theory together is now underway, and that’s what will be funded by that $1.58m.
The really big challenge comes in mastering spin: a materials issue, Kawakami said. “We need to get control over that process," he said. "We’ve done the best anyone can do, but it’s still not enough for this project."
The gate’s electrodes are ferromagnets. A small current is applied to the magnet that replicates the data as electron spins in the graphene. The idea, ultimately, is for the electrons in the graphene to mix and match the data to produce a logic operation, thereby achieving the goal of combined memory and logic operation on one chip.
Only, there are a couple of problems.
The first is how long a spin lasts: the best results are obtained the longer the spin lasts but right now the number of spins is well below the theoretical ceiling – by a factor of 1,000. Kawakami says the reason for this is unclear. “What’s limiting the spin is a big scientific mystery,” he says, noting that experiments are underway to improve the manufacture of graphene itself.
The other challenge is making the spin of the electrons – that's the process of copying the data from the electrode and putting it in the graphene – uniform. Right now, the process isn’t uniform or predictable enough for product, as some electrodes produce too much spin and others too little.
Kawakami’s team are piggybacking on evolving research in the field of magnetic RAM for answers and are developing new methods compatible with graphene. They are looking at spin torque, which controls the orientation of the magnetic electrode and thereby whether something is recorded as a zero or a one.
“We have to develop a special type of electrode geometry to get the spin torque to work with the graphene: that’s the main challenge,” he says.
This isn’t just about data transfer, however. The research is also considering the ramifications of power consumption – thereby determining the efficiency and green credentials of the finished magnetologic gate. “We think this is a really important part of the device because this will be the step that uses the most energy,“ Kawakami says. “So, we want to optimise this process as much as possible because this will be the energy cost of running the device.”
To further help control spin, a tunnel barrier is needed between the electrodes and graphene. The problem with spin is that it likes to leak out of the graphene, says Kawakami.
The spin doctor
The answer, it has been discovered, is to add a layer of insulation that keeps the spin inside the graphene, a discovery that saw the injection of half an atomic layer of titanium into the material. To achieve this, Kawakami’s team worked out how to grow graphene in sheets, something this material doesn’t like to do naturally.
“We put in a small amount of atoms – half an atomic layer of titanium – before we grow this film, and that prevents the molecules from forming balls,” Kawakami said. “I think there was a sense this was the important first step to make this material and it would work better.”
The benefit has been increased data retention, as the lifetime of spins has increased from 100 picoseconds to two nanoseconds.
Milestones planned for the next two years include the development of a completely new type of tunneling barrier and the demonstration of a new form of spin torque, writing magnetic bits that are compatible with graphene.
Long term, the team must work on the logic problem: how to assemble logic from different electrodes and spins. “We are doing experimental and laboratory work to understand basic processes," says Kawakami. "We have a fuzzy idea this could be good for logic and computation in general but that’s not our expertise.
“In our team, we have a concrete plan of a circuit we want to make. But that’s fairly rare... The general challenge is you have these gates that do these particular digital logic operations, or used to, and the challenge is how do you connect one device to another and transmit the other.
"Normally that’s done using volatile states. One of the general questions for doing spin logic is how do you connect different devices so an operation in one gate can be combined with logic operations on another gate to produce a higher level function?”
The first function the circuit will be built for is high-speed search.
Other challenges remain: can the graphene conduct the electrons fast enough while on power? Kawakami has just asked a colleague for deeper analysis.
Kawakami believes a working magnetologic gate can be delivered for use as a building block in other systems in three years’ time.
The reason is the project's remaining challenges are not show-stoppers. The real roadblock was blown away by the University of Manchester’s earlier work on graphene, which delivered the properties required for the construction of the gate and also on how to achieve spin at room temperature.
Now, it’s a matter of fine-tuning the graphene. When it comes to RAM, the team has options that include magnetic RAM, flash RAM and ferroelectric RAM. “There are many different candidates of other memory, It’s not a case of does it work but does it work better than other methods,” Kawakami said.
The University of Manchester has also been pushing the frontier on graphene in the area of tunnel barriers. “At a level where we can make this device – be able to make this device work - that part I’m very confident we can do,” Kawakami says.
He pauses, and re-phrases: “I've learned talking to engineers: there are more issues that come up that I hadn’t thought about. But at the level of can we make a gate, make it operate and operate well – definitely, we can do that.” ®