Less power-consuming graphics on your phone
Network-on-Chips are hardware that are used to communicate between different parts of a microchip. They have been around for over twenty years, but they still cannot be widely used in real life because they consume too much power. Computer scientist Peng Wang tried to find a solution for this power-inefficiency during his PhD. Defence on 12 February 2020.
A very promising solution
Manycore systems are systems that consist out of multiple cores. Each core has its own specific task. While a single core can process only one stream of data at a time, a manycore system makes it possible to process multiple streams of data. They are used in many technologies nowadays, for example for dealing with the graphics on your phone, your laptop and your game console. The more cores you add, the higher the performance of the manycore system. However, this leads to one main issue: the communication between all the cores becomes a bottleneck for the improvement of the performance.
A very promising solution to this problem is a Network-on-Chip (NoC), an intercommunication fabric for manycore systems that allows communication between the different cores of the system. However, these NoCs are not widely used in manycore system due to their high power consumption. PhD-student Peng Wang looked at two possible approaches to reduce power consumption in NoCs.
Dealing with high power consumption
‘In order to reduce power consumption, we can use power gating or we can use a simpler hardware design,’ Wang explains. ‘Network-on-Chips have a distributed structure, with a lot of different components. Not all components are necessary for every application of the NoCs. The first solution is therefore power gating, meaning that idle components of the NoCs are turned off to reduce power consumption.’ The current power gating approaches are inefficient and can increase latency. Therefore, Wang came up with three novel power gating approaches that reduce power consumption, without increasing the latency.
The second solution was using a simpler hardware design. Simple NoCs have a very low power consumption, but are unable to prioritize particular data streams or maintain a certain level of performance; requirements that are needed to use NoCs in manycore systems. Wang tried to find a way to use the simplified NoCs, but modify them so they could meet those requirements. He managed to do this by making a bufferless NoC, meaning that the NoC cannot temporarily store data. This makes the NoC less power-consuming, while still fulfilling all the requirements that are necessary to function in manycore systems.
Wang has made great improvements on the power consumption of NoCs. His final product is an NoC that is more power-efficient than traditional NoCs, making it possible to use them in manycore systems in the future. After finishing his PhD, Wang is planning to move back to China and find a job where he can combine research with engineering. His ultimate goal: finding a company that will manufacture his newly designed NoCs.