Design Space Exploration of Embedded SoC Architectures for Real-Time Optimal Control

Kris Shengjun Dong, Dima Nikiforov, Widyadewi Soedarmadji, Minh Nguyen, Christopher W. Fletcher, Yakun Sophia Shao

TL;DR

This paper tackles the challenge of running complex real-time robotic control algorithms, like Model Predictive Control (MPC), on power-constrained embedded systems. By exploring a wide range of hardware architectures (CPUs, vector processors, and specialized accelerators) and optimizing software, they achieve significant speedups and power reductions, paving the way for more capable and efficient robots.

What Problem Does This Paper Address?

Modern robots are becoming increasingly sophisticated, relying on complex algorithms for tasks like perception, planning, and control. Model Predictive Control (MPC), a popular technique for controlling dynamic robotic systems, is particularly computationally demanding. As robots shrink and become more power-conscious, fitting these demanding algorithms onto embedded System-on-Chips (SoCs) becomes a major hurdle. Traditional CPUs often lack the necessary performance or consume too much power, while specialized hardware can be difficult to program and less flexible. This paper addresses the critical need to understand and optimize the trade-offs between different hardware architectures and software strategies for running real-time robotic control on embedded platforms.

Methods

The authors conducted a comprehensive design space exploration, evaluating various hardware architectures:

  • General-Purpose CPUs: They tested different RISC-V CPU cores, from simple in-order designs to high-performance out-of-order superscalar cores (like BOOM).
  • Vector Machines: They focused on the RISC-V Vector Extension (RVV) architecture, specifically the Saturn-V processor, exploring different vector lengths and front-end configurations.
  • Domain-Specific Accelerators: They profiled Gemmini, a systolic array accelerator designed for deep learning, and explored its potential for robotic control tasks.

Their methodology involved:

  • Characterization and Benchmarking: Profiling kernel-level operations and end-to-end robotic scenarios (specifically, the TinyMPC algorithm) on these architectures.
  • Software Optimization: Developing optimized software mappings, including static scheduling, loop unrolling, operator fusion, and scratchpad memory management, to leverage the strengths of each architecture.
  • Hardware-in-the-Loop (HIL) Evaluation: Testing their optimized solutions on a real RISC-V vector SoC controlling a simulated drone, providing a realistic assessment of performance, power, and success rates.
  • Code Generation Flow: Proposing an automated code generation framework to simplify the process of mapping robotic workloads to specialized hardware.

Datasets

The paper does not explicitly mention training datasets. The primary "dataset" used for evaluation is the TinyMPC algorithm, a state-of-the-art embedded MPC algorithm. They also created simulated drone flight scenarios with varying difficulty levels (easy, medium, hard) involving waypoint tracking tasks.

Key Insights and Ideas

  • No One-Size-Fits-All: Different hardware architectures excel at different aspects of robotic control. Vector processors like Saturn show good general-purpose performance for MPC, while specialized accelerators like Gemmini can offer significant speedups for specific matrix-heavy operations but lack versatility.
  • Software is King (or Queen): Even with powerful hardware, aggressive software optimizations (like operator fusion and static scheduling) are crucial for unlocking performance gains, especially for tightly coupled, iterative algorithms like MPC.
  • Hardware Schedulers Aren't Always Helpful: For real-time control workloads, the complex scheduling mechanisms in vector cores and co-processors often provide minimal benefit and can even introduce overhead.
  • Vectorization Pays Off: Vectorized implementations of MPC can achieve substantial speedups (up to 3.71x) and significant power reductions (up to 27%) compared to scalar implementations.
  • Specialized Hardware for Specific Needs: For highly dynamic robots with high compute demands relative to actuation costs (like racing drones), high-performance SoCs are beneficial. For slower, more efficiency-focused robots (like hovering drones), lower-power, custom hardware might be more suitable.
  • The Engineering Bottleneck: Hand-optimizing software for specialized hardware is a significant challenge. The proposed code generation flow and the matlib library aim to alleviate this burden.

Key Results

  • Speedup: Achieved up to 3.71x speedups for MPC using optimized vector implementations.
  • Power Reduction: Demonstrated up to 27% system-level power reductions in robotic tasks.
  • Performance on Real Hardware: In HIL drone flight scenarios, vectorized MPC achieved near-ideal actuator power consumption and significantly improved disturbance rejection capabilities (e.g., enduring 1.89x larger forces).
  • CPU vs. Vector vs. Systolic: The paper provides detailed comparisons of performance, area, and utilization across these architectures, highlighting Pareto-optimal design points.
  • System-Level Impact: Accelerating MPC with vectorization freed up CPU resources, leading to a 1.35x improvement in the framerate of a concurrent CNN-based planning task (DroNet).

Limitations and Future Work

  • Hardware Availability: The HIL evaluation used a custom RISC-V vector SoC (Cygnus). While the paper provides detailed instructions, replicating the exact HIL setup might require access to this specific hardware.
  • Gemmini Versatility: While Gemmini shows promise, its current design is heavily tailored for DNN workloads. Future work could explore adapting it more effectively for general robotic control tasks.
  • Code Generation: The proposed code generation flow is a step towards automation, but further development could explore more sophisticated compiler techniques and representations for robotic workloads.
  • Heterogeneous Scheduling: The paper focuses on optimizing individual tasks. Future work could investigate strategies for scheduling multiple concurrent robotic tasks across heterogeneous embedded SoCs to maximize overall system performance and efficiency.
  • Future Hardware Enhancements: The authors suggest future architectural improvements for Gemmini, such as hardware GEMV support, which could further enhance its capabilities for robotic control.

Summary updated

Abstract

Empowering resource-limited robots to execute computationally intensive tasks such as locomotion and manipulation is challenging. This project provides a comprehensive design space exploration to determine optimal hardware computation architectures suitable for model-based control algorithms. We profile and optimize representative architectural designs across general-purpose scalar, vector processors, and specialized accelerators. Specifically, we compare CPUs, vector machines, and domain-specialized accelerators with kernel-level benchmarks and end-to-end representative robotic workloads. Our exploration provides a quantitative performance, area, and utilization comparison and analyzes the trade-offs between these representative distinct architectural designs. We demonstrate that architectural modifications, software, and system optimization can alleviate bottlenecks and enhance utilization. Finally, we propose a code generation flow to simplify the engineering work for mapping robotic workloads to specialized architectures.
BibTeX
@inproceedings{dong2025design,
  author = {Kris Shengjun Dong and Dima Nikiforov and Widyadewi Soedarmadji and Minh Nguyen and Christopher W. Fletcher and Yakun Sophia Shao},
  title = {{Design Space Exploration of Embedded SoC Architectures for Real-Time Optimal Control}},
  booktitle = {IISWC},
  year = {2025},
  doi = {10.1109/iiswc66894.2025.00047},
}
Figure