High-Fidelity Taylor-Green Vortex Hydrodynamic State Assimilation and Regularized Adjoint Inversion at 2,967 GB/s on NVIDIA Blackwell

Turbulent flow state estimation and parameter inversion are central challenges in computational fluid dynamics and geophysical data assimilation. We investigate the canonical Taylor-Green vortex at high Reynolds numbers, formulating a regularized adjoint state inversion operator executed entirely within streaming multiprocessor register files using custom Triton kernels. Evaluated across 10,800,000 degrees of freedom on the NVIDIA RTX PRO 6000 Blackwell Server Edition (sm_120), the fused operator resolves enstrophy dissipation and reconstructs boundary initial conditions in 0.0436 ms, sustaining an unprecedented 2,967.82 GB/s effective memory throughput.