Loading...
Technical Insight

Magazine Feature
This article was originally featured in the edition:
Issue 3 2024

Silicon-organic hybrid slot waveguide modulators on the verge of industrial adoption

News

Complementing highly scalable silicon photonics with engineered organic materials offers a path towards optical transceivers with unprecedented performance and efficiency. Recent data proves the photochemical stability of silicon-organic hybrid devices at practically relevant power levels, removing the last roadblock to industrial adoption.

BY Adrian Mertens, Carsten Eschenbaum, and Christian Koos AT SILORIX

Generative artificial intelligence (AI) is about to fundamentally transform our world, penetrating virtually every domain of modern life – from healthcare and diagnostics to industrial design and optimised production to stunningly realistic art, to name just a few. On a technical level, generative AI crucially relies on specific types of neural networks such as large language models (LLMs), which have tens of billions or even hundreds of billions of free parameters and which are designed to process and emulate human text or other content based on vast amounts of training data.

Training LLMs is a grand computational challenge, requiring massively parallel processing in dedicated AI clusters that contain thousands of highly specialised computing nodes, such as graphics processing units (GPUs) or tensor processing units (TPUs). However, scaling AI clusters – and hence the performance of associated LLMs – beyond current sizes is becoming increasingly difficult, primarily due to interconnect bottlenecks that limit the data transfer among the various computing nodes. Overcoming these bottlenecks is key to building even larger AI models that can take performance to yet another level.