Fujitsu develops clock distribution circuit

In recent years, in order to improve the data processing performance of servers, in addition to improving CPU performance, there has been a need to increase the speeds at which data is transmitted between chips, such as CPUs, in a server system. On the other hand, as speeds have increased, more power is consumed by the clock distribution circuits on the transceiver circuits that transmit the data. Now, by building a tiny oscillator circuit into each transceiver and synchronizing the oscillators, Fujitsu researchers succeeded in developing a low-power clock distribution method that eliminates the need to use conventional clock distribution circuits. It is expected that this technology will contribute to increasing the performance of the next generation of servers and supercomputers. Background In recent years, there has been a greater need for improvements in data processing performance for servers employed in datacenters that support cloud computing. This has led to enhancements in CPU performance, as well as the development of large-scale systems that connect large numbers of CPUs. As a result, the amount of data traffic exchanged between CPUs and peripheral devices has grown substantially. To accommodate this high volume of traffic, inter-processor data communications speeds in today's servers have increased from a few Gbps to tens of Gbps. In line with faster speeds, however, power consumption has also increased. For the next generation of high-performance servers, there is an increasing need to couple higher speed with low power consumption. Technological Challenges As inter-processor communications speeds have become faster between CPUs and other chips in server systems, the power consumed by transceiver circuits in transmitting the data has also increased. In particular, the proportion of the overall transceiver circuit's power consumption that is attributable to the clock distribution circuit has been growing. To make sure that the clock signal generated by the clock generator circuit is not attenuated and is able to be transmitted to each transceiver circuit, the clock distribution circuit consumes a considerable amount of power because of the relatively high signal amplification that is required over multiple stages. To achieve higher speeds with low power consumption, therefore, what has been needed is to reduce the power consumed by the transceiver circuit by developing a new, low-power clock distribution circuit. Newly Developed Technology By building a tiny oscillator circuit into each transceiver and synchronizing the oscillators, Fujitsu researchers succeeded in developing a low-power clock distribution method (figure 3) that eliminates the need to use conventional clock distribution circuits. All of the reference clocks (oscillator output) are set so that they are synchronized with the clock generator output. As a result, the frequency and phase of the reference clock for each transceiver circuit match the frequency and phase of the clock generator output. Because the signal between each oscillator required for synchronization used less amplification compared to conventional clock signal transmissions, less power is consumed. In achieving synchronization, a low amplification signal that is proportional to the amplification difference between neighboring oscillators is fed back into each oscillator (figure 4). Because this feedback is designed to drive the amplification difference to zero, ultimately the amplification difference becomes zero and the frequency and phase of the output clock for each oscillator become uniform. All neighboring oscillators are connected, with one connected to the clock generation circuit, meaning the operation of all oscillators is synchronized to the clock generation output. In this way, the generation and transmission of a low amplification differential signal enables the generation of multiple synchronized clocks. This technology enabled a 75% reduction in the power consumed by the 16 GHz clock distribution circuit, resulting in a 20% reduction in the overall power consumed by the transceiver circuit. Results It is anticipated that the application of this technology will make a significant contribution to improving the performance of the next generation of servers and supercomputers. Future Plans Fujitsu Laboratories and Fujitsu Laboratories of America will work to apply this technology to product areas related to big data, such as the backplane interfaces that link printed circuit boards together in servers.