All computer chips run on clocks. Different circuits carry out different pieces of work and calculations are completed in time with the chip’s clock. The next stage of work is not carried out until every calculation in the first stage is completed. The clock synchronises the functions of the chip, which means the chip can only work as fast as its slowest circuit. If chips continue to be made like this they could face functional problems as they get more powerful. Chips are getting physically larger with microprocessors having anything between three and six million transistors – and Digital Equipment Corp has one with nine million. The clock signal that regulates the the processing the chip is carrying out has to arrive at all parts of the chip at the same time.
Battery life
And the more transistors there are the more of a problem this becomes. In the future it may not be possible to get the clock cycle simultaneously around the larger chip because of the sheer number of transistors. There is also the power issue. Modern chips are usually fabricated in CMOS, which means that when the chip is not switching on or off it uses negligible power. If CMOS technology is used in a clocked chip, all the circuits switch each clock pulse even if they are not carrying out a calculation, so power is used all the time. The battery life of the increasing numbers of handheld computers and portable electronic devices means the issue of keeping power consumption to a minimum to prolong battery life is becoming more important. One possible solutions to this problem is asynchronous computing, or computing without a clock. Steve Furber, ICL Professor of Computer Engineering at the University of Manchester completed an asynchronous version of the ARM60 chip from Advanced RISC Machines Ltd in March this year. He argues that asynchronous computing is not inherently expensive and believes that the design, packing and fabrication of asynchronous chips is a straightforward process. The drawback is that although an asynchronous chip can be made, all current product-testing equipment is designed for checking clocked chips. However Furber said even checking asynchronous chips is an attainable goal. The collaboration between the University’s Department of Computer Science and Advanced RISC Machines Ltd announced this month formalises work that has been going on for years. The Cambridge-based company has bought the patent on techniques to fabricate the asynchronous chip that works without clocks.
By Abigail Waraker
Steve Furber and his research team at the University will work on the project to demonstrate the advantages of the aynchronous computing approach in the Amulet 2 chip, which should be completed by the middle of next year. The design software for the University’s original Amulet 1 chip was provided by Advanced RISC Machines, but now the firm is contributing funding as well. Funds will also come from the European Commission’s Esprit Open Microprocessor systems Initiative. Furber hopes to demonstrate an improved speed of up to three times that of the prototype Amulet 1 chip. He also expects to show better power efficiency, although at this stage he could not specify by how much. The design of the new chip is already under way. If Furber is right about asynchronous chips, it may help get around the clock skew problem. That is, the problem of getting the clock to all parts of the chip at the same time. The Amulet 2 will have no clock at all. All the signals will be generated locally and will actively make requests for service. In the simplest terms, the Amulet 2 chip will co-ordinate its internal operations by using additional circuits that will have the function of stating I am finished now and asking are you finished yet? Furber and his research team will work to develop this control circuitry, which will mean that instead of waiting for the next clock cycle to move onto the next calculation, a question and answer process takes place. The second calculation continues as soon as the first is finished. This means that the asynchronous chip
can work faster because it can work as fast as the speed of its average circuit rather than waiting for the slowest circuit every time. Advanced RISC Machines says the reason asynchronous computing has not taken off in the past is firstly the complexity of setting up the technology, but also that the additional control circuitry takes up space on the chip and adds to the expense. So, one of the tasks the University team faces is to develop more efficient control circuitry as well as improved design tools for creating the circuitry. Once the chip is developed, there is then the issue of integrating it into the normal clocked world. Advanced RISC Machines said it is not about to build a fully asynchronous chip because there would be compatibility problems. The computing world revolves around clocked chips.
In patches
So the asynchronous chip would have to interface to clock-based systems. Furber will be working on ways this can be done as part of the research. Advanced RISC Machines says the research has brought forward implications for the design of its clocked chips. It envisages using the technology internally in patches in its chips – although the company has not yet ascertained which functions within the chip will best be suited to asynchronous computing and is in the process of evaluating this at the moment. The ARM 8 family of chips is due out next year and although the chip will be fully clocked, it will make use of some of the techniques Furber has developed while creating the Amulet 1. Advanced RISC Machines sees the research as being on the leading edge and does not yet know how it will help develop further chips. There is the possibility that the chip that follows the ARM 8 might contain a wholly asynchronous chunk of circuitry, but there are no definite plans at this stage.
