Researchers have developed a new method to speed up the processing speed of electronic devices, such as smartphones and computers, without the need to replace existing components. This breakthrough could potentially double processing speeds while consuming less energy.
Modern devices are equipped with several types of electronic chips, including the Central Processing Unit (CPU), Graphics Processing Unit (GPU), hardware accelerators for artificial intelligence (AI), and digital signal processing units for audio. Usually, these components process data separately and sequentially, which can slow down overall processing.
To overcome this challenge, a team of scientists proposes an innovative framework where processing units work in parallel, rather than one after the other. This method, called "simultaneous and heterogeneous multi-threading" (SHMT), allows the different units to work on the same section of the computer code at the same time, significantly improving processing efficiency.
Unlike "software pipelining", a method that allows different components to work on different tasks simultaneously, SHMT allows for a more flexible task distribution among the components. This means that different processing units can tackle the same part of the code simultaneously, then move on to new tasks once their part is completed.
In addition to accelerating processing, SHMT proves to be more energy-efficient. Many tasks normally assigned to energy-hungry components, like the GPU, can be delegated to less energy-intensive hardware accelerators.
This approach was tested on a prototype system composed of a multi-core ARM CPU, an Nvidia GPU, and a TPU hardware accelerator, showing performance nearly twice as superior and a 51% reduction in energy consumption compared to a traditional system.
The adoption of this software framework on existing systems could not only reduce hardware costs but also decrease carbon emissions and the demand for fresh water necessary for cooling large data centers, thanks to a more efficient and ecological workload management.
However, the researchers note that their study, based on a prototype, requires further work to assess how this method can be applied in practical contexts and which types of applications could benefit the most.