CODESYS Control and a GPU Digital Twin on the Same Machine

Advantech is exhibiting at SEMICON Taiwan 2026 from 2 to 4 September at the Taipei Nangang Exhibition Center, in the Smart Manufacturing Pavilion, and the technical content of the stand is more interesting than the venue. The theme is real-time edge intelligence for semiconductor equipment, and the demonstrations describe an architecture that has been converging for a while.

The wafer fab equipment demonstration runs CODESYS-based EtherCAT control integrated with NVIDIA Omniverse libraries and Isaac Sim for digital twins and real-time control. That combination is worth pausing on. CODESYS on EtherCAT is deterministic machine control — cyclic, hard real-time, the thing that moves a stage accurately. Omniverse and Isaac Sim are GPU simulation. Putting them on the same equipment controller means the twin is not a model living in a data centre and consulted afterwards; it is running alongside the control loop.

The inspection and metrology demonstration uses high-speed CoaXPress image acquisition with the company's CamFlow vision software, and for advanced packaging there is a GPU-powered data acquisition system using GPUDirect zero-copy streaming with a DAQNavi SDK integrated with CUDA. Zero-copy is the detail that makes the rest viable: moving acquired data into GPU memory without a round trip through the CPU is what keeps a high-rate acquisition path from being throttled by its own plumbing. A SEMI S2-certified rack, SKYRack SEMI, covers the integration side, and autonomous mobile robots running real-time edge AI and an agentic 'AI Factory Brain' round out the stand. Linda Tsai, president of Advantech's Intelligent System Sector and chief operating officer, keynotes the Heterogeneous Integration Global Summit on 4 September.

Semiconductor equipment is where this kind of architecture appears first, because the tools are expensive enough to justify the engineering and precise enough to need it. But nothing in the pattern is specific to wafers. A deterministic control core, a GPU beside it for perception and simulation, and a zero-copy path between them is a general machine-building architecture — and the reason it is worth watching now is that the hard part has never been the GPU. It is keeping the real-time loop deterministic while something computationally hungry runs on the same silicon.