Verifiable Computation Under Embedded Constraints: Architectures, Energy Trade-offs, and Experimental Observation

Abstract

Verifiable computation has emerged as an important enabling technology for secure andauditable computing, yet its practical deployment on energy- and latency-constrained hardwareremains limited by the overheads associated with execution logging and verification. This articlereviews recent developments in transparent verification and zero-knowledge proof systems fromthe perspective of embedded and edge computing, with emphasis on physical constraints suchas power consumption, memory bandwidth, and real-time operation. As an experimental casestudy to ground the discussion, we examine an embedded witness-recording pipeline designed tosupport downstream STARK-style verification workflows.The evaluated system couples a delta-encoded witness-recording mechanism with cryptographic integrity binding and was implemented on a Raspberry Pi 5 equipped with a Hailo-8neural processing unit. Under the reported test configuration, the recorder sustained approximately 11.47 million recorded steps per second, with per-step recording latencies on the order oftens of nanoseconds. The associated computation kernel—a Lyapunov-stable linear contractionoperator defined in PyTorch, exported via ONNX, and compiled to Hailo-8 silicon—exhibitedmeasured execution rates of 49,859 operations per second in hardware-only mode and 28,394operations per second when invoked through a Python-mediated runtime.Witness data were captured using Cython-optimized routines with delta encoding andmulticore sharding, producing approximately 1.4 GB of compressed shard data for 500 millionrecorded steps, corresponding to an observed compression ratio of roughly 14× under the testedworkload. Integrity binding was achieved via cryptographic commitment roots validated throughrecomputation and contract-level checks, with integration demonstrated using Cairo smartcontracts on StarkNet. Throughput-derived frame-budget estimates, along with energy andcost calculations based on stated power-accounting assumptions, are presented as contextualindicators rather than general performance bounds.All implementation artifacts, execution logs, and validation results are publicly released tosupport independent inspection and reproduction. The reported measurements characterize thebehavior of a specific witness-recording architecture under a constrained workload and hardwareconfiguration, and are intended to inform broader discussions of verifiable computation forembedded systems rather than to establish general claims about end-to-end zero-knowledge proofgeneration performance Baramay Station Research Inc.Canadian Non-Profit R&D (Saskatchewan)Director & Principal Investigator: Michael Tass MacDonald Contact [email protected]