Accelerator Polymorphism: Transcending Domain-Specific Architectures with Robotics

Robotics is poised for a ChatGPT-like leap, but progress is constrained by the lack of a compute substrate suited to its cross-domain computational demands. Unlike LLMs and neural networks, which are largely confined to a single domain of tensor operations, robotics spans multiple algorithmic domains, including graph search, control theory, optimization, and more, often within a single application pipeline. This cross-domain diversity, coupled with stringent power, latency, and cost constraints, demands a new design point-one that transcends the siloed boundaries of domain-specific accelerators without reintroducing the overheads of general-purpose processing. To address this challenge, we propose a novel architectural concept: accelerator polymorphism, whereby a single physical substrate dynamically shape-shifts itself from one specialized execution model to another without sacrificing the efficiency of specialization by reverting to general-purpose design. We instantiate this concept in a concrete microarchitecture, Morphatron, that dynamically shape-shifts across five execution morphas-queuecentric SIMD, graph, tree, vector, systolic-each tailored to distinct algorithmic domains in robotics. At the heart of Morphatron is Morpha Core, a novel SIMD processor specialized for queuecentric execution and built on a key algorithmic insight: queues are a unifying data structure across robotic workloads. Morpha Core elevates queues to first-class SIMD operands, integrates memory allocation directly into the pipeline as an architectural primitive, and thereby enables hardware acceleration of dynamically-sized data structures. Morphatron is a multi-core accelerator composed of a two-dimensional array of Morpha Cores, designed to support the diverse access and parallelism patterns revealed by our indepth study of end-to-end robotic workloads. We evaluate the performance and power efficiency of Morphatron using six end-to-end robotic applications from the RoWild benchmarks [1]. Compared to the NVIDIA Jetson Orin Nano GPU, Morphatron achieves a 5.5× speedup and a 6.6× improvement in performance-per-watt, on average. Relative to the ARM A78 CPU, Morphatron delivers a 7.7× speedup and a 7.7× improvement in performance-per-watt.