A world of profoundly distributed computing is upon us, if researchers at the University of California at Berkeley are to be believed. They are developing 'smart dust' — sensor-laden networked computer nodes that are just cubic millimetres in volume.
As they develop the various elements of smart dust, researchers are also creating new units of measurement, where data transfer and computing performance will be measured in nano or even picoJoules of energy consumed per bit handled.
The smart dust project envisions a complete sensor network node, including power supply, processor, sensor and communications mechanisms, in a single cubic millimetre. At these dimensions, the nodes would be small and cheap enough to be scattered from aircraft for battlefield monitoring, stirred into house paint to create the ultimate home sensor network or swallowed with your breakfast cereal for health monitoring.
The Berkeley project is chipping away at the problems it needs to overcome in order to make smart dust a reality. Throughout its work, it is focusing on reducing the amount of energy used by the dust motes by every means possible. That means custom processors, minimal operating systems, handcrafted memory designs, special sensors and converters and novel communications systems.
Professor Kris Pister, who leads the research, has laid out some of the limits his team is working with in a discussion paper on smart dust. For example, acquiring sample data from a sensor takes around 1nJ. Sending a bit of data over 10 to 100m by RF takes 100nJ. But sending the same bit 10m over a collimated light beam will take 10pJ per bit, making it 10000 times more efficient than RF.
Similarly, a 32bit computation in a power-optimised microprocessor takes around 1nJ per instruction processed, while the engineering limits suggest this could be cut to 1pJ per instruction with dedicated hardware.
On the other side of the equation, Prof Pister says batteries can store about 1J of energy per cubic millimetre. Solar cells can provide approximately 100µW/sq cm in bright sunlight and more than 100nW in the same area in average room lighting. Devices that scavenge energy from ambient vibrations can gather nano-watts per cubic milli-metre.
Putting all these estimates together, Prof Pister reckons that the energy to acquire and process a sample and then transmit some data about it could be as small as a few nanoJoules. Given that a cubic millimetre battery can store 1J and could be backed up with a solar cell or vibrational energy source, smart dust motes could run for years.
So what does smart dust consist of? The team has already experimented with macro and micro-motes, large-scale demonstrator projects using commercially available devices to test the concepts. It is now working on a series of circuit and micro-electro-mechanical systems (MEMS) designs to cast those functions into custom silicon.
Brett Warneke is a project researcher who, along with handling the design of a custom low-power processor, is also working on the integration of the aspects of smart dust into a working system. He is currently finishing the design of an ultra-low energy processor with a new instruction set, tuned for distributed sensor networks. Along with the energy-saving design of the core, Warneke has included autonomous subsystems to ease control of the A/D converter, receiver and transmitter.
The optical receiver for the smart dust project is being developed by Brian Leibowitz, a student at the University of California. The receiver senses incoming laser transmissions at up to 1Mbit/s, for a power consumption of 12µW. Although this is too high for continuous use in smart dust, it is a reasonable figure for the download of small amounts of data such as a 1Kbit program.
For data transmission, the team is exploring a passive system which draws on the ancient Greek concept of the heliograph — a mirror or highly polished surface that glints in the sunlight to send messages across long distances. For smart dust, the team is using corner cube retro-reflectors (CCRs) built using MEMS techniques. CCRs are produced by placing three mirrors at right angles to each other to form the corner of a box that has been silvered inside.
The key property of a CCR is that light entering it is reflected back along the path it entered on. For the smart dust system, the CCR is being built on a MEMS process with the two vertical sides being assembled by hand. When a light is shone into the CCR, it reflects back to the sending position. By modulating the position of one of the mirrors, the reflected beam can be modulated, producing a low-energy passive transmission.
Lixia Zhou is developing the CCR for the project. So far, she says the device has shown good mechanical characteristics including a flat mirror surface and good mirror alignment. Its functional performance has also been good, with actuation at 5V and a data transmission power consumption of about 50pJ/bit.
Zhou has tested the CCR with a mirror side length of 300µW, illuminated with a 0.8mW laser. With this set-up, the CCR successfully transmitted meaningful data with a high signal-to-noise ratio over 180m in air. She points out that the transmission distance could be much greater with larger mirrors and a more powerful illuminating laser. During the feasibility phase of the project, the team sent data 21km across San Francisco Bay using a modified laser pointer whose flashing was detected by a video camera connected to a laptop.
The fourth major part of this smart dust demonstrator is the analog-digital convertor (ADC). Mike Scott, who is developing the 8bit ADC, has so far demonstrated an input range of 1V, equal to the power supply, and a 70kHz sampling rate. The converter draws 1.8µW when sampling at that rate, or 27pJ for an 8bit sample. Scott claims a simple modification can bring this down to 10pJ/sample and that the ADC's off power consumption is too small to measure.
In a further bid for energy efficiency, the converter has been designed with a successive approximation architecture so that it can deliver any number of bits of accuracy up to the maximum eight. This means that in situations such as thresholding, further energy can be saved.
Scott says this is particularly useful in distributed sensing networks where you may only need approximate values. He believes that with a cubic millimetre battery, his ADC could take 10sample/s for more than a century.
The team is looking at a two-chip solution which separates the control circuitry from the MEMS and sensor arrays. One chip will have the micro-controller, memory, ADC, optical receiver and a light sensor. The MEMS chip will carry the CCR, a solar cell, an accelerometer and its drive circuits. All this integration should be completed within the next six months. The team has shown the CMOS integration, but without the memory and microcontroller. Each MEMS part works but has yet to be integrated.
There are further issues — for example, getting the solar cell and the optical receiver to work together since they need special optical filtering materials to work side by side.
According to Warneke, the latest smart dust mote, with a volume of just 16cu mm, has been tested. It takes samples from a photo-detector, transmits their values with the CCR and runs off solar cells. However an onboard accelerometer is not yet working and the optical receiver needs the filters before it can work alongside the solar cells.
So smart dust is on the way. The team has a number of the elements they need to make up an intelligent sensing node. And other sensors and communications mechanisms, such as short range RF, are also in the works.