Forwarding Is Not Free: Measuring The Performance– Area Trade-Off Of Data-Hazard Handling On A Minimal RV32I Pipeline
IOSR Journal of VLSI and Signal processing · 2026-04 · 16 (4) · 12-26
DOI: 10.9790/4200-16041226 ↗Kurz gefasst
Fachartikel in der Zeitschrift IOSR Journal of VLSI and Signal processing, 2026-04. Verfasst von Pranav Vinaya Kumar.
Abstract (Originalsprache)
Pipelining overlaps the execution of successive instructions to approach a throughput of one instruction per cycle, but the data hazards it creates force a designer to choose between stalling, which is simple but wastes cycles, and forwarding, which recovers those cycles at the cost of extra hardware. This trade-off is taught in every undergraduate architecture course, yet students rarely see it measured on a single controlled implementation where forwarding is the only thing that changes. We built two RV32I five-stage cores that are identical in ISA, datapath, memory model, and control-hazard handling, and differ only in their data-hazard logic: Core A stalls until the producing instruction retires, while Core B adds an EX/MEM and MEM/WB forwarding network with a one-cycle load-use interlock. Running four benchmarks that stress hazards differently through a cycle-accurate simulator, we find that forwarding lowers cycles per instruction (CPI) by 49.7% on a dependency-heavy kernel and 26.6% on a memoryheavy kernel, but by exactly 0.0% on a control-dominated kernel that contains no close data dependencies for it to resolve. On a mixed bubble-sort workload — our headline case — CPI falls 12.0%. A synthesislevel area and timing model puts the cost of the forwarding network at 11.96% more lookup tables and a 7.4% reduction in maximum clock frequency. Once that frequency penalty is folded in, forwarding still wins in wall-clock terms on three of four benchmarks, but is actually 8.0% slower on the control-heavy case, where it buys no cycles yet still lengthens the critical path. The overall picture: on this core, forwarding is worth building — but not universally, and not for free.
Zitieren
APA
Kumar, P. V. (2026). Forwarding Is Not Free: Measuring The Performance– Area Trade-Off Of Data-Hazard Handling On A Minimal RV32I Pipeline. IOSR Journal of VLSI and Signal processing, 16(4), 12-26. https://doi.org/10.9790/4200-16041226
Vancouver
Kumar PV. Forwarding Is Not Free: Measuring The Performance– Area Trade-Off Of Data-Hazard Handling On A Minimal RV32I Pipeline. IOSR Journal of VLSI and Signal processing. 2026;16(4):12-26. doi:10.9790/4200-16041226
BibTeX
@article{Kumar2026,
author = {Kumar, Pranav Vinaya},
title = {Forwarding Is Not Free: Measuring The Performance– Area Trade-Off Of Data-Hazard Handling On A Minimal RV32I Pipeline},
journal = {IOSR Journal of VLSI and Signal processing},
year = {2026},
volume = {16},
number = {4},
pages = {12--26},
doi = {10.9790/4200-16041226}
}
EUVIMED