German semiconductor industry eyes ASICs and chiplets as GPUs hit energy limits

Graphics Processing Units (GPUs) have dominated high‑performance computing for over a decade, serving both graphics and machine‑learning workloads. Their parallel architecture, however, encounters growing energy and cost constraints when powering large AI models. Data‑center operators in Germany and across Europe confront rising electricity prices and stricter efficiency standards, prompting a shift toward more specialized hardware.

Application‑Specific Integrated Circuits (ASICs), modular chiplets and analog compute units are projected to gain prominence by 2026. ASICs, such as Tensor‑Processing Units, deliver far higher energy efficiency for dedicated tasks by eliminating unnecessary functions, though they risk rapid obsolescence if algorithms evolve. Chiplets enable flexible, scalable designs, while analog accelerators offer low‑power alternatives for certain neural‑network operations. The slowdown of Moore’s law and the physical limits of transistor miniaturisation are steering innovation from smaller transistors to new architectures.

A recent analysis of 500 GPU models from 2009 shows power draw rising from under 200 W to more than 600 W for the latest high‑end cards, turning powerful gaming rigs into substantial heat sources and increasing cooling demands. These trends underline the industry’s move away from universal GPUs toward task‑optimized silicon solutions.