IBM Labs has unveiled a new electronic technology that integrates on the same silicon layer both optical and electrical features, paving the way for the creation of faster and more efficient computers.
The new technology developed by IBM is called CMOS Integrated Silicon Nanophotonics (CISN), and provides the main benefit of integrating on the same chip the traditional technology used in computers (the electronic technology) with the optical one, mainly used in telecommunication systems for broadband transmission (ie on fiber optic). This technology promises an important and innovative application: build a computer (or rather, a supercomputer) in which communication between the various units may be produced by light pulses (optical pulses) instead of through the normal electrical signals. What are the benefits that follow? Essentially three:
- it will be possible to get faster devices
- it will be possible to build devices with more compact size
- device will have a more efficient power management with respect to those that can be realized today with the traditional technology
The new technology is the result of over 10 years of work done by IBM's global research labs, and is expected to achieve a chip integration density up to 10 times than current. CISN technology will pave the way for the construction of Exascale computing systems, that is essentially supercomputers able to perform a million trillion floating-point math operations per second (1 exaflop). An exascale supercomputer would be able to achieve a level of performance even 1000 times higher than best performant existing machines.
Besides combining optical and electrical components on the same silicon layer, the new technology can also be used in conventional CMOS fabrication lines, and does not require a new or particular production process. It follows that future silicon transistors will be able to share the same silicon layer with silicon nanophotonic devices. In order to enable this important innovation, IBM researchers have already developed a suite of silicon nanophotonic devices, integrated and ultra compact, both active and passive.
These components are extremely small, with a size scaled down to the
diffraction limit, the smallest size that dielectric optics can reach. By slightly changing the current standard CMOS manufacturing process, it will be possible to create a multitude of silicon nanophotonic devices integrated with high-performance silicon CMOS devices, either analog than digital, such as: optical modulators, germanium photodetectors, and ultra-compact frequency-division multiplexers. The following image shows a layer made with the revolutionary technology developed by IBM.
This second image, instead, shows the team of scientists who participated in the research group on CISN technology: from left to right Yurii Vlasov (Director of the Silicon Nanophotonics Department of IBM Research), William Green, Solomon Assefa.
The ability to produce devices with CISN technology using existing production facilities for the traditional silicon CMOS technology is a huge advantage, both in terms of production costs, both because it makes the new technology usable in record time (they are already talking about production in the current year, 2011). The density of integration, both optical and electric, is very high and unprecedented: a single channel transceiver, complete with its optical and electrical circuitry, occupies only 0.5mm2. It will also be possible to create individual transceiver with an area of only 4x4mm2, and capable of receiving and transmitting data at speeds of several terabit per second (trillion bits per second).
The results of research conducted by IBM has been communicated by Prof. Yurii Vlasov at the SEMICON conference (the largest international conference for the semiconductor industry), which was held in Tokyo in December 2010 (the original title is "CMOS Integrated Silicon Nanophotonics: Enabling Technology for Exascale Computational Systems").
Optics means Terabit per second
The current trend in high performance computing systems is to increase the parallelism in the implementation of the code through an extensive use of multithreads, increasing the number of processors or blades, or ultimately increasing the number of cores per processor. The scale effect that follows is leading processing speeds higher and higher toward the fateful exaflop threshold. However, there remains the problem of the available bandwidth required to make these exaflop travel within the system, bandwidth which in itself is limited.
Optics is just to fill that bandwidth limiting, bringing with it a huge bandwidth which derives from the use of high modulation speed and from the parallelism of typical WDM (Wavelength Division Multiplexing) systems. In other words, it is a kind of extension at the microscopic level of what happened years ago at the macroscopic level to broadband trasmission systems (the "backbones"), where traditional copper connections were replaced by fiber optic connections.