In an era where we are witnessing the adoption of applications artificial intelligence As it spread widely across various sectors, we must strive to find more efficient and effective ways to consume energy. In recent years, data centers that rely more and more on artificial intelligence technologies have witnessed a sharp rise in energy consumption rates, to the point that they have become a significant contributor to global energy consumption and emissions rates.
According to a report issued by McKinsey in September 2024, data centers in the United States of America (excluding cryptocurrency centers) are expected to witness a three-fold growth in their energy needs by the end of this decade, and that we will also witness a similar rise in energy demand. In other regions of the world.
However, AI-based solutions may hold the key to addressing today’s biggest dilemmas. How can we continue to develop and use powerful AI-based models while simultaneously contributing to building sustainable, zero-carbon economies?
Mohammed bin Zayed University of Artificial Intelligence takes a multi-pronged approach to the research it conducts to address this global challenge.
Researchers at the university focus on multiple research axes, including graphics processing units and tensor processing units. Traditional computing structures, such as central processing units, are not efficient enough to operate artificial intelligence technologies, which has led to the development of graphics processing units and tensor processing units as specialized tools designed to meet the requirements of operating artificial intelligence technologies and carrying out the resulting processing operations. It should be noted that graphics processing units and tensor processing units form the cornerstone of artificial intelligence. Current data centers use tens of thousands of them. Therefore, operating the new generation of data centers currently being built will require larger numbers of these units. In order to achieve the continuity of the artificial intelligence revolution, it is necessary to work on developing these devices in ways that ensure energy savings and improve the performance of their processors with the lowest possible energy.
Hence the importance of designing devices dedicated to artificial intelligence (such as graphics processing units or tensor processing units) that are capable of saving energy, as this would contribute to enhancing the efficiency of data centers in energy consumption, especially centers that manage increasing numbers of complex and complex calculations. With the aim of developing individual components and harmonizing their designs to reduce power consumption at the hardware level without affecting software performance.
The deeper we look into ways to improve hardware architecture, the more solutions we will find that will address this problem from its roots.
The smallest unit that carries or transmits information in all computing operations is known as a “bit,” which is a number in the binary counting system that has two possible values: 1 or 0. Finding a way to conserve the energy consumed to store, retrieve, and process the values 1 and 0 in operations Computing reduces the waste of energy and heat, speeds up processing processes, and at the same time enhances energy efficiency. Removing “unnecessary” bits in expensive computation (through quantization or pruning) goes a long way in addressing energy efficiency issues. It should be noted that the challenge lies in redesigning the devices and enhancing their efficiency without compromising their computational capabilities and accuracy.
As for transistor technologies, which are constantly evolving and becoming smaller and smaller in size, their reliability must be guaranteed to avoid any errors. Hardware errors have long been a problem for large companies, such as Google and Meta, that run large data centers. Building reliable processors will address energy efficiency issues and help build a sustainable future. The longer the processors are used in data centers, that is, the more reliable they are, the lower their carbon footprint, because building these processors requires huge initial costs.
In addition, FPGAs are another alternative solution that provides customizable devices designed to perform specific AI-based tasks. This matrix helps developers optimize the electrical circuits in devices for different applications, which contributes to reducing the energy they consume and maintaining their high performance.
It must be noted that the Mohammed bin Zayed University for Artificial Intelligence is also looking for ways to reduce waste, distribute resources more efficiently in the upper classes, and enhance the sustainability of the processes of developing artificial intelligence models and their applications.
For example, the system software used in the training and inference phase of large language models must be closely aligned with the hardware design, in order to enhance its energy efficiency.
The research conducted by researchers is divided into: Mohammed bin Zayed University for Artificial Intelligence Currently, the aim of achieving the sustainability of artificial intelligence systems is divided into two parts. The first part is to optimize and distribute AI model training tasks through dense nested operators using heterogeneous resources in the GPU server, including for example the use of compute-intensive matrix multiplication with network-intensive communication operations, from In order to improve systems performance and use available resources more efficiently.
The second part is to work to limit the computing operations that occur in the inference stage of artificial intelligence models, by publishing models that are adjustable, smaller in size, more specialized, and more effective, and by storing the answers (or intermediate results in the inference process that led to those answers). ) in the model temporarily in order to benefit from it in processing similar tasks.
Ultimately, the more AI technologies develop, the greater the need for solutions that reduce energy use. Ambitious organizations in the UAE can benefit from specialized computing architectures and also adopt software with innovative designs in order to significantly reduce the energy use rates required by artificial intelligence applications, which contributes to reducing their environmental impact and at the same time enhances their economic feasibility and the possibility of widespread adoption. Broader, and thus contributes to achieving sustainability goals within the framework of the country’s future vision.
