Quantifying the Memory Wall and Ensuring Reproducibility in a Legacy Spectral GCM: A Case Study in Thread-Safe OpenMP Modernization

Legacy spectral Global Circulation Models (GCMs) face severe challenges on contemporary multicore and wide-vector architectures due to static memory paradigms and mixed-precision bottlenecks. We present a systematic refactoring of the Main Geophysical Observatory (MGO) spectral GCM (T42L25) that achieves OpenMP thread-safety with strict IEEE-754 bitwise reproducibility across 1-16 threads. Migration of COMMON blocks and SAVE variables to a heap-based THREADPRIVATE architecture, a deterministic per-J term-split reduction for spectral synthesis (eliminating undocumented inter-latitude state leakage via VE_JRESET), and uniform REAL*8 promotion across the radiation module together unlock clean AVX-512 auto-vectorization. On an AMD Ryzen 9 9950X, the modernization yields a 4.73× total speedup at 16 threads relative to the legacy sequential baseline, with OpenMP strong scaling of 4.36×, and saturates DDR5 bandwidth at the STREAM Triad ceiling. Cross-platform validation on a 2013-era dual-socket Intel Xeon E5-2697v2 (Ivy Bridge-EP, DDR3) confirms the memory-wall signature: a 2.25× CPU-frequency increase yields a near-zero median wall-time effect (0.35%), direct empirical evidence of bandwidth-bound execution. Bitwise-identical outputs are reproduced across 1-24 threads on both platforms. Peak Energy-to-Solution reduction reaches 49.0% at the energy-optimal thread count. Parallelizing the currently-serial CVDDRI convection-initialization is identified as the natural next refactoring target.