The Dresden conference centre was designed to represent a stark modern contrast to the restored Baroque buildings of the old town of Dresden. For some reason, the architects decided to build a curved building with one floor on a slope, cutting through other, flat, floors. The entrance is up a long flight of stairs, exposed to the wind and rain blowing across the river Elbe. The conference rooms are all provided with wonderful glass walls overlooking the river, which have to be blacked out if you want to be able to see the information projected onto the screen.
However, it is the the main conference venue for Silicon Saxony, a “cluster” of high tech companies ranging from semiconductor manufacturers Global Foundries and Infineon to a wide range of supporting and related businesses: over 300 companies are part of the network. And it is where, in alternate years, the academic community engaged in EDA make their spring pilgrimage to the DATE Conference. DATE (Design Automation and Test in Europe) calls itself The European System Design Show: from Systems–on-Chip to Embedded Computing. (However the embedded computing discussed here had little overlap with the recent embedded world exhibition in Nuremberg). This year DATE had six tracks of deeply technical papers from PhD and post-doc students and their supervisors, an exhibition theatre with presentations of slightly less deep material, and an executive overview / special day topic thread that included the conference keynotes. These papers tended to be a series of overviews of different aspects of the current scene.
Looking for things that would interest you, dear reader, I found that the best sessions were from the executive/special day thread that were looking at some of the new technologies that might allow Moore’s Law to continue.
Some of the topics included were organic semiconductors, magnetic technologies including spintronics, on-chip optical techniques, and new technologies for memories.
Before we come to these, there was an interesting session on Dark Silicon. The name was coined by Michael Taylor of the University of California, San Diego, where he has established a Center for Dark Silicon. His argument is that, as we have continued to shrink transistors, we have not been able to reduce their power demands in proportion. Straight shrinking and running at high frequencies would have chip surface temperatures at close to the exhaust temperature of a jet engine. One way around this is to split processing across multiple cores, but even here, the power demands are such that cores are run at lower frequencies than their capability, or cores are powered down for greater or lesser periods of time, depending on the application running. The same approach is also used for areas of logic in a large chip.
Taylor calls the effect of these approaches “Dark Silicon”. He is at pains to point out that dark silicon is not useless silicon, just underused silicon. The answer, he feels, is not to shrink further or to look at better power-down approaches or to wait for what he calls a “deus ex machina”, some sort of hitherto unexplored approach, such as new transistor architectures or new material. Instead he proposes an approach that uses Conservation Cores (C-Cores): specialized cores, each of them tuned for the task at hand (10-100x more energy efficient), and turn on only the ones that are needed. These are automatically generated co-processors, created in the system design flow. In use, program execution jumps between the host CPU and the relevant processor(s). The Dark Silicon Center has been working on their GreenDroid mobile application processor: a tile design tuned to an Android workload with nine C-Cores and a host processor. Simulation has shown massive energy savings, and a physical prototype with 4 tiles is being evaluated.
Another approach to the heat problem is the idea of liquid cooling. A paper from IBM and EPFL (the École Polytechnique Fédérale de Lausanne) looked at both surface cooling and the use of micro-fluidics to cool chips in a 3D stack. With the micro-fluidics approach, a mixture of liquids is pumped through micro-channels in the stack. By choosing the right chemicals and using the walls of the micro-channels as electrodes, the fluid can generate enough power to run the chip as well as cooling it. An analogy is with the blood of a mammal, which acts both as heating/cooling and nutrition.
Now to some specific technologies: with transistors shrinking but energy use increasing, a major factor in energy consumption is the power burned by the interconnect. Optical interconnect offers to decrease the power needed while increasing the connectivity speed. Optical is already in use in high-speed backplanes for servers and supercomputers, but it is still at very early stages for on-chip use. A paper presenting the work of a Franco/Canadian/American collaboration looked at techniques for replacing silicon electron paths with optical interconnects. The Chameleon project is working on optimising the network structure for connecting multiple IP cores on a large-scale SoC, using a separate layer of a 3D stack for the light paths. Initial results suggest that a fairly simple network can provide both flexibility and speed in routing.
When I first heard the phrase “Organic semiconductors”, my mind went to the fruit and vegetable displays at the supermarket, not to the multiple volumes of chemistry books with Dewey classification 547 that I shelved when working in libraries. Carbon is already attracting attention in microelectronics, as graphene, a one-atom thick layer of carbon, has some interesting electron properties. But when you start building molecules with both the carbon atom and atoms of other substances, things start to get interesting. The nineteenth century scientists exploited this to create dyestuffs and pharmaceuticals, including aspirin. The petrochemical industry’s numerous by-products were developed with organic chemistry. There is enormous knowledge about how to synthesise large molecules of carbon with other materials, and this is being used to custom-build organic semiconductors.
One of the leading researchers in this area is Karl Leo of the Technical University of Dresden. His analysis is of four waves of organic semiconductors. The first wave is using OLEDs (Organic Light-Emitting Diodes) for displays, the second using OLEDs as lighting, the third building Solar cells, and the fourth for integrated circuits.
The advantages of OLEDs include: they are a larger light source than the pin-point of a conventional LED, they don’t need back-lighting like other flat-screen displays, and they can be sandwiched to stack RGB sources rather than have them clustered. The issues that are still being resolved are how to increase the basic efficiency of the devices and also how to improve the transmission of the light from the organic layer through the other layers that provide the electrodes and the viewing screen.
These same issues are holding back the wide take-up of OLEDs for lighting, even though the potential is enormous – the light sources can be curved, and they can be a mix of colours for specialist applications.
The flexibility and light weight of organic materials also make them very attractive for photo-voltaic applications, but, again, there is still a need to improve efficiency.
Using organic materials to create integrated circuits was the subject of another paper, which reported work on printing circuits onto flexible plastic substrates. There was a quick look at displays that are flexible enough to be rolled up tight and stored in a relatively small device yet they can be unrolled to create quite a large screen. But the main thrust was in building circuits. These are already being built with one-micron channel lengths for p-channel TFTs. (A demonstration of a DAC, with a 6-bit stream, is already in a laboratory.) There is work going on that will see other variants of the organic materials providing n-channel devices.
A team that includes Germans, Swiss, and Americans, is working on creating materials that have even greater electron mobility. It wasn’t entirely clear, or I didn’t understand, but I think they are building simulations of molecules that should yield the required behaviour and then scanning the literature to see if that configuration has been created or can easily be created.
Alternative memory technologies were the subject of another session. An overview by Roberto Bez of Micron looked at the vast field of today’s memory market. Data centres soak up vast quantities of memory, as do personal devices and desktop machines. And most ASICs, SoCs, processors, and FPGAs also have a considerable area devoted to memory. While there are increases in the use of non-volatile memories, DRAM and SRAM still have massive market shares, and research is looking at ways to improve densities and reduce power consumption. With DRAM, one route is to start looking in three dimensions. On-chip approaches include using vertical capacitors and transistors. Another approach is to have 3D chip stacks, with extra memory occupying its own layer. And non-volatile NAND is being created with on-chip stacks 32 bits deep. But none of these approaches is as yet keeping up with the demand for greater density and improved read-write performance.
Matthias Wuttig of Aachen University of Technology called phase-change memory “pretty cool stuff”. And, while we are using phase-change routinely in our desktop machines as the technology for optical storage on re-writeable DVDs, when you look at using it to replace flash memory, it becomes complicated, as well as cool. Phase-change materials can be rapidly switched between amorphous and crystalline states, usually by being heated and cooled rapidly. Heating and cooling is fine when you have a laser and relatively unlimited space, but it is much more difficult on-chip. There is a wide range of potential phase-change materials, and the search is on to understand the wider properties that these materials possess, but it is probable that they will provide non-volatile memories with the switching speed of DRAMs.
Another area where full understanding of the underlying mechanism that causes an effect is yet to be reached is that of resistive memory (ReRAM). In a ReRAM cell, a layer of resistive material becomes conductive through the application of voltage. This is usually because the voltage causes a filament of conducting material to grow through the insulator. This can be reversed by applying a different voltage. Normally the insulator is one layer in a sandwich, which is built from an electrode, the insulator, a layer of conducting material to contribute the filament, and a second electrode. The main research is looking at the best materials for the two layers of sandwich filling that will form filaments at relatively low power consumption and will not have power leakage.
Printing on thin films is not restricted to organic semiconductors. Thin Film Electronics, based in Norway and Sweden, sells ferro-electric memory printed at high volume on reel-to-reel printing presses to create very low-cost electronic labels. These can be written to and read from quickly and easily in the field. But the company is now moving towards more complex applications, with sensors and logic, and they showed a temperature-sensing and display label that is cheap enough to go onto individual items for one-time use. This can record whether food and other products being shipped have been exposed to temperatures outside predetermined ranges. Today the standard approach is to use tags that monitor entire containers. Since DATE, Thin Film Electronics has announced a deal that allows PakSense, a company selling the container-level monitors, to sell tags that can be applied to pallets and individual items.
The final area that is interesting here is spintronics. The angular momentum of an electron can be sensed as though it were a small magnet. Spintronics can change that angular momentum, effectively reversing the magnet and so creating a zero and a one. Using spintronics has allowed greatly increased density of hard disk drives, which is why it is now commonplace to have a terabyte of storage on the desktop. Several companies are working to use spintronics for solid-state memory and also for logic. As with other technologies, the search is on for materials that will make spintronics fast, low power, and cheap.
It is possible that not all these different technologies will succeed, but the innovation that they could bring will be what keeps the electronics industry continuing its drive forward.