Marian Verhelst
Papers
7
Total Citations
61
H-Index
5
About
Marian Verhelst is a hardware architect and embedded systems researcher whose work sits at the intersection of energy-efficient computing, irregular graph processing, and precision-scalable arithmetic. Best known for pioneering the DAG Processing Unit (DPU) series, Verhelst has tackled one of the most stubborn challenges in modern computing: efficiently executing irregular directed acyclic graphs that arise in probabilistic machine learning, sparse linear algebra, and robotic navigation — domains where conventional CPUs and GPUs struggle with unpredictable data dependencies. The DPU and its successor DPU-v2, alongside the stream-based PIU processor, demonstrate how custom silicon architectures can deliver dramatic energy efficiency gains while supporting novel number formats like posit arithmetic. Verhelst's contributions extend beyond graph processing. Early work on dynamic sensor-frontend tuning addressed always-on embedded classification under tight power budgets, while more recent research into Microscaling (MX) data types targets on-device robot learning at the edge. The GraphOpt framework further showcases a systematic, optimization-driven approach to parallelizing irregular workloads. Collectively garnering citations across systems, circuits, and machine learning communities, Verhelst's research charts a clear trajectory: making intelligent, adaptive computation practical for resource-constrained real-world systems.
Research Focus
Key Achievements
Top Papers
- 1
- 2
- 3DPU-v2: Energy-efficient execution of irregular directed acyclic graphs10 citations · 2022
- 4
- 5
- 6
- 7