Comparative Evaluation Methodologies for Algorithm Acceleration on Embedded Platforms: A Structured Review and Reproducible Assessment Framework

Background: Performance results for algorithms accelerated on microcontrollers, field-programmable gate arrays (FPGAs), multiprocessor systems-on-chip (MPSoCs), embedded GPUs, and application-specific accelerators are often reported under different workloads, accuracy requirements, and measurement conditions. Objective: This review examines how published studies measure execution time, throughput, energy, memory use, and hardware resources, and proposes a consistent approach to interpreting comparisons across platforms. Methods: A targeted review of publisher records, research repositories, and benchmarking standards identified 15 representative implementations published between 2015 and 2026. The analysis draws on 52 references, including primary studies, surveys, and methodological sources. Study outcomes were recorded with their associated tasks, platforms, measurement conditions, and level of source verification. Because the search was not an exhaustive database-level systematic search, the article is presented as a structured review rather than a PRISMA-compliant systematic review. Results: Reported outcomes vary substantially in both units and experimental scope. One FPGA CNN implementation reported 61.62 GFLOPS at 100 MHz; FINN reported 21,906 image classifications per second under a stated system power of less than 25 W; and a 2026 microcontroller study reported 38 ms execution time, 387 µJ per inference, and 174 kB of SRAM in a specified configuration. These results cannot be ranked directly without matching workload and measurement conditions. Discussion: Differences in power boundaries, memory definitions, hardware configurations, and baseline optimization limit cross-study interpretation. The proposed Cross-Platform Accelerator Evaluation Protocol (CAEP) provides a sequence of checks for workload equivalence, measurement boundaries, and reporting completeness before numerical comparisons are made. Conclusion: A defensible comparison begins by establishing whether results measure equivalent tasks under sufficiently similar conditions. The proposed protocol requires prospective validation, and a comprehensive systematic review would require a fully documented database search and independent full-text assessment.