The brain as an inspiration for future processors

The brain as an inspiration for future processors

The human brain is inspiring new processors that may perform cognitive functions

The brain as an inspiration for future processors

Antonio González, researcher from the Department of Computer Architecture at the UPC, who has been awarded the Advanced Grant

Groundbreaking knowledge

The “CoCoUnit:An Energy-Efficient Processing Unit for Cognitive Computing” project has received an Advanced Grant, the highest award granted by the European Research Council (ERC) to groundbreaking research projects.

It is scheduled to start in September and will go on for five years. It has received 2.5 million euros in funding, which will be mainly allocated to hiring research and research support staff, as well as to equipment and expenses related to the research project (publications, attending conferences, etc.).

Researching the design of new intelligent computing systems inspired by the human brain is the goal of the “CoCoUnit: An Energy-Efficient Processing Unit for Cognitive Computing” project, led by Antonio González, a researcher from the Department of Computer Architecture at the UPC. He has received an Advanced Grant, the highest award granted by the European Research Council to investigators pursuing groundbreaking high-risk projects.

May 21, 2019

The next revolution in computing that we will witness in coming years will be driven by deploying smart devices around us in all kinds of environments—work, home, entertainment, transport and health care, for example—, backed up by smart servers in the cloud”, explained the researcher Antonio González. There is a growing interest in extending the capabilities of computing systems to perform human-like tasks, such as speech recognition, language translation, speech synthesis, image classification and object recognition. That is, in equipping computers with learning, synthesis and reasoning abilities equivalent to those performed by the human brain (what is known as cognitive computing).

These cognitive computing systems will provide new user experiences by delivering new services or improving the operational efficiency of existing services, such as self-driving cars, so that they can process, like we do when we are driving, images and objects (traffic lights, pedestrians, acoustic signals such as ambulance sirens) on the road in real time and, in the field of health care, systems for monitoring patients 24/7.

A key feature of these cognitive computing systems will be their capacity to process in real time large amounts of data from audio and video devices and other types of sensors. This will demand very high computing power but at the same time extremely low energy consumption”, added González. The energy-efficiency requirement is an essential condition for success, not only for mobile and wearable systems but also for large data processing centres whose energy consumption is very high and a main component of the total cost of operation.

This is the starting point of the research project “CoCoUnit: An Energy-Efficient Processing Unit for Cognitive Computing”, which has been recognised with an Advanced Grant, the highest award granted by the European Research Council (ERC) to research projects. Current processors are not a good fit for developing this type of task since they keep the same basic structure as early computers, which were mainly optimised for number crunching.

Learning from neuroscientists
The project is based on a disruptive approach: investigating unconventional, “dramatically different”—in the words of the researcher—architectures that can offer a greater energy efficiency in terms of performance and cost for cognitive functions. “Our brain is the most efficient system currently known for tasks such as image recognition and language processing. We aim to design new architectures of computer systems based on how the brain works that perform cognitive functions with a high level of energy efficiency and at a reduced cost, so that they can be integrated into all types of devices”, explained González.

Thus, the goal of the project is to devise a new processing unit, the CoCoUnit, that will be integrated with the existing units of a processor (general-purpose cores and GPUs) and will be able to perform cognitive functions with extremely high energy efficiency. “We envision a new processing unit that will equip future processors (in a way similar to that in which most processors today include a graphics processing unit) and deliver new user experiences that require cognitive functions to be performed in real time”.

The idea is to learn from neuroscientists’ observations on how the brain works and how neural networks are structured to improve processors. Despite all the information it processes, our brain is very energy-efficient”, explained González, who is the director of the research group Microarchitecture and Compilers (ARCO) and a researcher at the Department of Computer Architecture at the Universitat Politècnica de Catalunya · BarcelonaTech (UPC). Thus, the future CoCoUnit will be based on a massively parallel architecture with extremely simple units. A large number of simple units are known to be more energy-efficient than fewer more complex units with similar performance.

CoCoUnit, the new processing unit
Furthermore, the new unit will reduce data movement. The von Neumann architecture used by processors today involves huge power consumption for moving data around the system: every instruction and operand has to be fetched from memory and sent to the execution units, and results have to be written back to the memory hierarchy. The interconnections for moving these data consume the majority of energy of a microprocessor. Therefore, reducing data movement can provide important benefits in energy efficiency.

The CoCoUnit will include specialised hardware for some key functions of these applications. Specialisation is a key method for improving energy efficiency, but it increases the production cost, since specialised units can only be used for a restricted set of applications. A key challenge is to identify the most important functions for specialisation. It will also rely on a different computing model, oriented to “intelligence”: learning, rather than “imperative programming”, will play a key role in this new approach. In addition, the new unit will exploit resiliency and approximate computing for even greater energy efficiency.