Answers are grounded in this paper and cite their sources.
Embedded and edge devices increasingly require lifetime adaptability: once deployed, they must accommodate evolving workloads, standards and protocols, and dependability and security requirements under strict energy and cost constraints. This paper presents a runtime-reconfigurable FPGA-based RISC-V microcontroller architecture that combines dynamic partial reconfiguration with standardized custom-instruction integration via SCAIE-V. Accelerators are implemented as reconfigurable modules hosted in reconfigurable partitions and invoked via a custom instruction. We showcase the approach with two soft cores – the non-pipelined PicoRV32 and the pipelined Orca – showing that the design is mainly independent of the underlying microarchitecture. Beyond standardized accelerator integration, we contribute a system-level execution mechanism that preserves ISA-level semantics under runtime reconfiguration. When an accelerator is absent, dedicated hardware transparently triggers a controlled context switch to an equivalent software fallback function, then restores the architectural state so that execution proceeds as if the custom instruction had executed in hardware, including program-flow changes. To maintain correctness for stateful accelerators during reconfiguration, we introduce a memory handover manager that enables seamless migration across the hardware/software boundary. An automatic reconfiguration module loads accelerators on demand while the software fallback remains available. We evaluate the concept on an AMD Artix-7 FPGA using a control-flow-intensive reduced-overhead loop and an ASCON-based lightweight cryptographic accelerator. For ASCON, hardware execution reduces runtime from 126,017 to 3515 clock cycles on PicoRV32 and from 66,191 to 1800 clock cycles on Orca, while the software fallback adds only 20.23% and 18.04% overhead over regular software, respectively. In contrast, the ROL case study exposes the fixed cost of full semantic transparency, reaching 1289.93% and 1089.45% fallback overhead. These results show that the approach is most beneficial for accelerators with sufficient per-invocation work and reuse to amortize fallback and reconfiguration costs, while maintaining correctness and software-level compatibility under dynamic accelerator availability.
No takes yet. Share an insight, caveat, or question.
Scheipel et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: