From breakthroughs in generative models to debates over ethics, regulation, and its role in society, AI is on the front page daily. But amid the buzz, we often forget to talk about the engine under the hood: the fundamental compute technology that makes it all possible.
That conversation needs to shift. Because behind every AI innovation lies a pressing reality - innovation and new approaches will forever evolve in terms of capability, novelty and optimisation, and the hardware we choose today will define AI’s capabilities tomorrow.
The 'Great Shift' in computing
If you take a step back, what we’re witnessing isn’t just progress in AI. It’s a foundational shift in computer science. For decades, computing has been rooted in sequential logic. Deterministic, rule-based, step-by-step programming. But AI doesn’t fit that mould. We’re moving into the age of probabilistic programming and computing, where models are trained, not hard-coded, and behaviour emerges from data and probability, not predefined logic.
This shift changes everything. And it’s forcing us to rethink how we design hardware.
In the early days of neural networks, purpose-built accelerators like Neural Processing Units (NPUs) and other customised processor approaches emerged to meet the specific demands of known AI workloads (e.g. CNNs). They were powerful, but rigid. As AI evolved, the workloads became more dynamic. New architectures, operators, and approaches are arriving faster than specialised hardware can adapt.
The result? These largely rigid, fixed-function accelerators are increasingly unfit for purpose.
AI is a parallel compute problem
Strip away the jargon, and at its core, AI is a parallel compute challenge. AI models - especially deep learning networks - rely heavily on matrix operations, vector arithmetic, and large-scale parallel processing. More importantly, processing most AI models can be looked at as a huge number of smaller, semi-autonomous calculations. In that sense, it’s not so different from rendering graphics.
In fact, the same challenges that led to the rise of the GPU in gaming and graphics - rendering millions of pixels in real time - are re-emerging in AI.
The solution then was general-purpose, massively parallel computing. The solution now is the same.
AI doesn’t need another tightly optimised, inflexible chip. It needs a scalable, adaptable, parallel compute platform. It needs a GPU.
GPUs: The underappreciated AI platform
While the market has been conditioned to think of NPUs as the default for AI, that mindset is due for an overhaul. Generalised parallel compute is not just sufficient for AI - it’s ideal. And the GPU is the most mature, flexible parallel compute platform available.
Companies like Imagination have long been building GPUs capable of this. As compute shaders and general-purpose GPU (GPGPU) capabilities advanced, architectures evolved well beyond traditional graphics. Today, with products like the IMG E-Series, Imagination offers a GPU IP family purpose-built for the edge - blending programmable graphics and AI acceleration in one cohesive platform.
That’s a game-changer. It’s not just about raw performance; it’s about flexibility. A general-purpose GPU can evolve with AI itself, adapting to new workloads and approaches without needing a new chip every cycle.
Changing the conversation
There’s growing support across the industry for this view, but we’re still in the early days of awareness. The market hasn’t fully realised that AI’s future depends less on highly specialised silicon and more on scalable, general-purpose architectures.
We need to re-educate the ecosystem: the GPU is no longer just for rendering - it’s the cornerstone of the next era of computing.
If we want AI to scale, to reach the edge, to power everything from cloud-based models to smart cameras and autonomous vehicles, we need an efficient platform built for general-purpose parallel compute.
We need the GPU.
Author details: Dennis Laudick, Vice President of Product Management, Imagination Technologies