World model planning accelerator reduces robot power and heat use
The World Model Planning Accelerator (WMPA): Hardware-Algorithm Co-Design for Embedded AI in Robotics
Hardware ArchitectureMachine LearningRobotics
Summary
Robots often predict what might happen next before making a move, but doing these predictions needs a lot of energy and makes hardware very hot. The authors analyze the fundamental energy limits of these prediction steps and find that current GPUs use far more energy than necessary and cause high temperatures. They design a new chip architecture called the World Model Planning Accelerator that uses smarter memory and computation techniques to reduce power use and heat. This makes running predictions directly on robots much more practical.
What this means in practice
- •For robotics hardware designers: Build robot chips that efficiently perform predictive planning at low power and heat using near-memory compute and adaptive precision techniques.
- •For mobile device engineers: Develop energy-efficient embedded AI accelerators for real-time prediction tasks under strict power and thermal constraints.
Authors
Prerana Ramkumar, Suhas Suresh Bharadwaj, Fazal Hameed
Abstract
Embedded AI is moving from perception to prediction. Robots now use world models$\unicode{x2014}$learned simulators that predict the outcome of candidate actions in a compressed latent space before any is executed. Every predicted step must be computed on board within a tight power and thermal budget, yet the energy these steps fundamentally require, and how far real hardware sits from it, have not been characterized. We derive lower bounds on the arithmetic operations and the CMOS dynamic energy of a latent state transition from information theory and Boolean circuit complexity. Profiling a DreamerV3-style Recurrent State Space Model (RSSM) shows that every rollout operation is memory-bandwidth bound and that a GPU consumes at least $5.9\times10^{10}$ times the CMOS energy floor. Transient 3D thermal simulations show that this inefficiency is also a thermal one, since sustained back-to-back planning heats a millimeter-scale silicon tile to 556 K even behind an ideal copper heat spreader. Guided by this chain from bound to bottleneck, we propose the World Model Planning Accelerator (WMPA), a chip architecture for embedded world model planning that combines near-memory computation, real-time entropy estimation, variable-precision rollout units, and copy-on-write latent scratchpads. With a design target of 40x lower power, the same tile stays near ambient at 306 K in simulation, pointing toward practical planning on thermally constrained robots.