Embedded physical artificial intelligence (AI) combines sensing, learned inference and physical action under resource and timing constraints. This focused review synthesizes 62 detailed primary-source records across seven workload discussions and eight operation groups into conditional hardware requirements. Six cross-paper comparisons yield five adaptation judgments and one conditional functional match; none validates concurrent physical sharing. The reviewed audio fast Fourier transform (FFT) engine supports maximum lengths of 512/1024/2048 points at 32/16/8 bits per real or imaginary component, whereas the radar engine reaches 4096 points using half-precision arithmetic. Further configuration-specific contrasts expose memory and scheduling tradeoffs. An analytical MCUNetV2 example reduces peak activation storage from 1372 to 172 kB with 10% more multiply–accumulate operations. Published physical inverter tests at 40 kHz report parallel/sequential calculation times of 50/400 ns with 0.7% current total harmonic distortion in both; the corresponding synthesized designs use 200/70 digital signal processing cells. These findings motivate joint selection of numerical format, storage, execution granularity and timing budgets. A quantitative requirements matrix, explicit compatibility criteria and a conceptual architecture connect the evidence to configurable arithmetic, banked storage, state management, arbitration and software interfaces, with RISC-V as an implementation target. The framework distinguishes reusable operations and designs from concurrent resource sharing and specifies the numerical, task-quality and timing tests required to validate proposed mappings.