Against the backdrop of edge computing converging with large models, the chip industry is undergoing a profound architectural shift. For a long time, to handle diverse workloads such as image rendering, speech recognition, computer vision, and neural-network inference, chip vendors have tended to keep stacking specialized accelerators into SoCs—CPU, GPU, DSP, and NPU each doing their own job.
But that approach creates its own problem. “Every accelerator is very good at solving one problem, but every accelerator also creates a dependency.”
At Imagination Technologies’ recent media open day, CEO Markus Mosen described that cost as a “border tax.” Different compute units behave like isolated islands. Data must constantly travel through external memory among GPU, NPU, DSP, and CPU. Every time it crosses a boundary, three costs are paid at once: energy, latency, and engineers’ development and maintenance time.
Imagination wants to do the opposite: subtract. It wants the GPU to evolve from a pure graphics processor into a fusion accelerator spanning graphics, general-purpose compute, and AI. The company recently published compute-performance figures for its latest E-Series GPU IP. The substantial gains further reinforce its GPU-first strategy. The E-Series uses one processor, one architecture, and one software stack to handle graphics, compute, and AI workloads together.
The “border tax” of fragmented architectures, and the rise of the GPU as the center of gravity
From the end user’s point of view, a phone, a car, a robot, or an industrial camera does not treat “perception, modeling, inference, decision-making, and rendering” as five unrelated tasks. They sit on the same data path, face the same response-time budget, share the same battery, and share the same memory and power envelope.
Read more
