Open Access
  • 19 November 2026
    • Vol. 15  (22) , 5390
    • DOI:10.3390/electronics15225390
Abstract
Single-event upset (SEU) hardening of native D flip-flops (DFFs) in flash-based field-programmable gate arrays (FPGAs) lacks unified experimental conditions and quantitative selection criteria; this work implements six DFF structures on one and the same A3PE3000 chip as identical dual 243-bit shift-register test chains. These include an unhardened baseline DFF (BL-DFF), a conventional triple-modular-redundancy DFF (TMR-DFF), three delay-filtering TMR-DFF variants (DF-TMR-DFF), and one dual-path combinational-logic-replica TMR-DFF (DCLR-TMR-DFF). Realized DF-TMR-DFF filtering windows are set by on-chip process–voltage–temperature (PVT) calibration rather than nominal design values. Six heavy-ion species are used, with linear energy transfer (LET) of 4.4–82.06 MeV·cm2/mg, and irradiation data are processed through a five-stage procedure to remove terminal output-stage single-event transients (SETs) and retain genuine bit-cell upsets. All cross-sections are normalized per logic bit (one complete instance) rather than per storage cell. At the highest LET, the DCLR-TMR-DFF achieves the lowest cross-section of 8.23 × 10−10 cm2/bit, approximately 1/145 and 1/27 of the BL-DFF and TMR-DFF cross-sections, respectively; the DF-TMR-DFF cross-sections decrease from 1.58 × 10−8 to 2.06 × 10−9 cm2/bit as the filtering windows widen. The proposed analytical model shows that which of the DF-TMR-DFF and the DCLR-TMR-DFF has the lower per-bit VersaTile overhead depends on the delay-stage count relative to the combinational-to-sequential resource ratio ρ0, with measured per-bit overheads from 11 (BL-DFF) to 37 (DCLR-TMR-DFF) VersaTiles. Post-layout static timing analysis (STA) of standalone projects gives maximum frequencies of 59.034, 52.384, and 43.092 MHz for BL-DFF, TMR-DFF, and DCLR-TMR-DFF, and 40.827–32.083 MHz for the DF-TMR-DFF variants, whose radiation-hardness gain comes at the expense of timing performance. No structure is optimal across all three dimensions, and trade-offs are required in design. A four-tier engineering selection boundary is established for 130 nm ProASIC3 space circuits. All TMR-based structures share an inherent single-point vulnerability at the terminal three-input majority voter.

This publication has 28 references indexed in Scilit: