AI research

AI evolves faster than the hardware it runs on. Frontier models land on hardware designed years before their workload existed. Meanwhile, next-generation chips and systems are developed without a clear view on the algorithms they will actually need to support.

Imec closes this loop — co-designing next-generation AI with next-generation hardware, from silicon to systems, and from research to application. 

Imec's AI activities are built on two complementary pillars:

1. Full-stack hardware-software co-optimization — applying and co-designing AI across the full stack: from workloads and system architectures down to chip design, semiconductor research, and fab operations.
2. Next-generation AI paradigms (imec.AI-labs) — exploring novel AI model architectures, guiding hardware requirements beyond today's paradigms.

Together, they form a continuum from long-term AI innovation to applied solutions, and deployment at scale.

Full-stack hardware-software optimization

Imec’s AI work combines deep expertise in semiconductor technology with advanced knowledge of algorithms and system architectures. Our guiding principle: AI systems must be co-optimized across algorithms, architectures, and technology. We model and optimize the impact of today's and tomorrow's AI systems onto today's and tomorrow's hardware.

Our partners can benefit from this work through a comprehensive set of initiatives that allow them to evaluate and optimize how next-gen AI systems translate to next-gen hardware:

• SiliconAI to access an authoritative source on how today's AI systems drive TCO.
• EdgeLab to test out their own algorithms and hardware in the edge.
• Imec.kelis to model the performance of future silicon.

SiliconAI

SiliconAI is a publicly available benchmarking initiative that identifies how today's AI workloads influence bottlenecks on the hardware and drive total cost of ownership (TCO) at the system level. It offers unique value to:

• AI model builders who want to test how their novel model architectures perform against real silicon systems.
• Infrastructure providers looking to make future-proof hardware investments.
• Hardware providers aiming to showcase their silicon’s capabilities on a vendor-neutral benchmark.

Visit the SiliconAI website to find out more and explore a partnership

EdgeLab

EdgeLab is an EU-funded community and platform focused on accelerating edge AI benchmarking. It allows developers to benchmark their model on the latest commercial and state-of-the-art edge devices in minutes.

Industry teams can confidently move from AI development to production-ready edge solutions––reducing risks, accelerating iterations, and lowering development costs.

Start benchmarking on the EdgeLab website

Imec.kelis

Imec.kelis is a performance modeling and design space exploration tool for LLM data centers.

Consisting of an LLM task-graph analyzer, a parallelism mapper, a hierarchical roofline model, and a topology-aware collective communication library, it offers an end-to-end framework to quickly and accurately evaluate and optimize your design choices.

Visit the imec.kelis webpage to download the spec sheet and contact us for a collaboration

AI for sensing and actuation across application domains

Imec’s work across a broad range of application domains gives it insight into domain-specific needs and challenges, and how AI can address them. For example:

• In health, AI-driven insights and predictions can improve biosensor design and biomanufacturing flows, or significantly increase our knowledge in domains such as neuroscience or proteomics.
• Advanced ADAS and autonomous vehicles are redefining the automotive industry and rely heavily on AI that is tightly co-designed with various sensing modalities.
• The advent of physical AI announces a new era for robotics, introducing autonomous systems that flexibly interact with the physical world and with other robots or humans.
• Smart, AI-enabled sensor networks allow for the monitoring of critical infrastructure or personnel in the domain of security and resilience.

Imec co-designs AI across the full sensing-to-actuation stack — from sensor and simulation to edge AI and autonomous actuation.

Imec.AI-labs — Next-generation AI paradigms

Imec.AI-labs explores future AI approaches beyond the current state of the art, acting as imec’s long-term innovation compass.

Central to imec.AI-labs' approach is the conviction that the next leap in AI will not come from scaling a single architecture, but from orchestrating diverse, specialized agents that reason, collaborate, and adapt — much like heterogeneous compute systems themselves.

The research is structured around three core directions:

• Language as a universal coordination protocol. Natural language serves as the interface through which agents exchange observations, intentions, and constraints — removing the need for rigid, task-specific APIs. This enables zero-shot integration: new agents (software services, hardware controllers, or human operators) can join a system without bespoke interface engineering.
• Anticipatory intelligence. Instead of merely reacting to past data, imec.AI-labs develops models that reason about possible futures — answering questions like "What would happen if this action were taken?" — enabling agents to evaluate alternatives, assess risks, and make proactive decisions before committing to action.
• Continuous adaptation and knowledge integration. A dual-memory architecture combines fast-updating symbolic memory with long-term parametric learning, allowing agent societies to incorporate new information without catastrophic forgetting — essential in non-stationary environments such as semiconductor research and clinical practice.

By developing these multi-agent societies in tight co-design with novel hardware primitives, imec.AI-labs generates insights that flow back into imec's semiconductor roadmap while simultaneously informing the broader AI community.

In hardware/software co-design, agent societies treat hardware constraints, software requirements, and application-level objectives as elements of a shared dialogue, exploring design spaces more efficiently and producing solutions tailored to specific workload–silicon pairings.

In healthcare, multi-agent systems model interactions between patients, clinicians, and institutions — integrating medical knowledge with patient-specific data to evaluate interventions and support personalized clinical decision-making.

Research is conducted through an open collaboration model: short, intensive bootcamp programs bring together imec researchers, academic partners, and industry collaborators around focused challenges. Results are published, prototyped, and — where they mature — handed over to imec's applied teams for integration into customer-facing solutions, ensuring a direct pipeline from frontier research to real-world impact.

Discover imec.AI-labs's research