Leading the charge: GPUs powering the explosive AI data centre growth

As nations across the globe ramp up investments in AI data centers and cloud computing, the term "Sovereign AI" is gaining popularity. These centres, often referred to as "AI Factories", are creating a surge in demand for highly specialised computing chips, particularly GPUs (Graphics Processing Units). A report by IDTechEx highlights the trajectory of AI chips, forecasting significant growth in the next decade.

GPUs are becoming indispensable, capturing a whopping 82% of the AI chip revenue in 2024. By 2025, their deployment is expected to multiply, dominated by industry leader NVIDIA with its Blackwell GPUs. Close on its heels, AMD competes fiercely with its MI300 and MI350 series, securing substantial deals with major technology companies.

Initially developed in the 1970s for basic 2D graphics rendering, GPUs have undergone significant transformations. The 1990s witnessed a growth in 3D graphics, with AMD and NVIDIA developing technologies that allowed GPUs to harness parallel processing capabilities for broader uses, such as simulations and image processing by the mid-2000s.

The surge of interest in AI in the 2010s, propelled by models like AlexNet and ResNet, further cemented the role of GPUs in training advanced AI models. Modern-day GPUs are tasked with facilitating complex AI operations, ensuring high-speed processing and supporting vast library functions needed for deep learning.

Comprised of thousands of cores, each GPU is designed to execute specific instructions simultaneously across numerous data points. Despite their simpler cache systems compared to CPUs, GPUs enhance throughput efficiency, crucial for tasks involving extensive data calculations.

The future will likely see high-performance GPUs adopt advanced transistor nodes, such as 2nm, a move that promises greater efficiency and density. However, challenges persist, particularly with the considerable costs of ultra-advanced lithography equipment and other hurdles, such as increasing heat production and materials limitations.

While custom ASICs and emerging chip technologies challenge the GPU stronghold, GPUs remain dominant, thanks to technological innovations like die-stitching and chiplet 3D stacking. Such innovations increase transistor counts and improve yield rates, though often at the cost of memory speed.

High-bandwidth memory technologies, led by Samsung, SK Hynix, and Micron, are widely adopted. This ensures the necessary memory to train expansive AI models, with Chinese enterprises now entering the HBM production arena. As this industry continues evolving, GPUs are poised to play a central role in shaping the future of AI data centres.