Driving · 2026-11-10

The World Model Planning Accelerator (WMPA): Hardware-Algorithm Co-Design for Embedded AI in Robotics

Prerana Ramkumar, Suhas Suresh Bharadwaj, Fazal Hameed

arXiv:2611.10316PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors analyze the computational and physical cost of embedded world-model planning. They derive information-theoretic and CMOS-based lower bounds for latent rollout steps, then profile a DreamerV3-style RSSM on a GPU and run transient 3D thermal simulations. They report that GPU rollout energy is ~11 orders above a CMOS dynamic energy floor and that sustained planning heats a mm-scale tile to 556 K. They propose WMPA, an accelerator architecture targeting 40× lower power.

Why it matters

World-model planners (e.g., MPC/MCTS) require many sequential latent rollouts, which must run within tight power and thermal budgets on edge robots. The paper connects theoretical minimum energy bounds to measured hardware inefficiency and then to thermal feasibility, motivating a hardware-algorithm co-design tailored to sequential rollout structure, memory bandwidth, and precision/entropy variability.

Method

  • Derive lower bounds on arithmetic operations and a CMOS dynamic energy floor for a latent state transition using information theory, circuit complexity, and Landauer’s principle assumptions for overwriting entropy.
  • Empirically profile a DreamerV3-style RSSM rollout on an NVIDIA Tesla T4 using a roofline model and sustained-load energy telemetry (pynvml) to locate the energy gap’s cause.
  • Run transient 3D non-linear thermal simulations under sustained back-to-back planning to test whether rollout power density drives a mm-scale silicon tile to unsustainable temperatures; then propose WMPA with NMCU/EEH/VPRU/LSS.

Limitation

The authors state that (i) the precision schedule rests on a synthetic entropy-collapse proxy, (ii) the 40× reduction is a design target and not a synthesis result, and (iii) the thermal model places an ideal sink behind the spreader.

Abstract (from arXiv)

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.

Related papers