Skip to content

GPU liquid cold-plate three-level multi-objective optimization (1D → FVM → LBM)

English | Japanese

« Optimization, inverse problems and uncertainty · « Home

Simultaneous three-level multi-objective optimization of a staggered cylindrical pin-fin liquid cold plate cooling a 700 W-class GPU module, running 1D screening, Grace CPU steady FVM optimization, and Hopper H100 unsteady 3D thermal LBM Pareto verification concurrently on a single Miyabi-G (GH200) node in 13.7 minutes.

Unsteady 3D thermal LBM of a staggered pin-fin GPU liquid cold plate (velocity magnitude and temperature, 3.07 M cells)

1D to FVM Pareto front to unsteady thermal LBM verification and Miyabi-G GH200 simultaneous CPU/GPU utilization

Problem setup

  • Domain: Periodic channel slice \(L = 36\text{ mm} \times W = 12\text{ mm} \times H = 3.0\text{ mm}\), heated die zone \(x \in [6, 30]\text{ mm}\) at \(q'' = 437.5\text{ kW/m}^2\) (equivalent to 700 W over a 40 mm × 40 mm die; \(Q_{\text{slice}} = 126.0\text{ W}\)).
  • Fluid & solid: Water at \(T_{\text{in}} = 25\,^\circ\text{C}\) (\(\rho = 994\text{ kg/m}^3\), \(\nu = 7.2\times 10^{-7}\text{ m}^2/\text{s}\), \(\text{Pr} = 4.82\)), copper pin fins (\(k_{\text{Cu}} = 385\text{ W/(m}\cdot\text{K)}\)).
  • Design variables (4): Pin diameter \(D \in [1.0, 2.4]\text{ mm}\), streamwise pitch ratio \(S_L/D \in [1.6, 3.2]\), transverse pitch ratio \(S_T/D \in [1.6, 3.2]\), inlet velocity \(U_{\text{in}} \in [0.18, 0.65]\text{ m/s}\).
  • Objectives (both minimized): Pressure drop \(\Delta P\) [Pa] and chip peak temperature \(T_{\max}\) [°C].

Simultaneous resource allocation on 1 Miyabi-G (GH200) node

Level Method & grid Hardware Throughput Mean utilization
1D model Zhukauskas / Short staggered pin-fin correlations + fin efficiency \(\eta_f\) Grace CPU (1 core) 2,000 candidates / 0.1 s (8 Pareto seeds) —
Steady FVM blockMesh + snappyHexMesh (140k–190k cells) + buoyantBoussinesqSimpleFoam (\(k\)-\(\omega\) SST) + Optuna NSGA-II Grace CPU 64 cores (4 workers × 16 MPI) 116 trials (~20–26 s/trial), 15 Pareto designs 71.9% (82.9% active)
Unsteady thermal LBM D3Q19 TRT+Smagorinsky LES + D3Q7 thermal LBM (nvc++ -stdpar=gpu), 480×160×40 = 3.07 M cells × 70,000 steps Hopper H100 GPU + 8 Grace cores 15 Pareto verifications (55 s each, 3,933 MLUPS mean) 96.4% (~390 W)

Multi-fidelity comparison across Pareto designs

Design (D / S_L/D / S_T/D / U_in) 1D \(\Delta P\) [Pa] FVM \(\Delta P\) [Pa] LBM \(\Delta P\) [Pa] 1D \(T_{\max}\) [°C] FVM \(T_{\text{avg}}\) [°C] LBM \(T_{\max}\) [°C] LBM \(T_{\text{rms}}\) [K]
Low-ΔP (1.84 mm / 3.12 / 2.77 / 0.32 m/s) 168 191 184 71.1 79.8 77.4 1.92
Knee (1.02 mm / 2.42 / 1.99 / 0.25 m/s) 988 464 715 46.6 60.0 52.7 0.57
High-cooling (1.10 mm / 1.77 / 2.90 / 0.54 m/s) 5,318 1,148 2,173 36.8 64.6 40.3 0.23
Min-T (1.08 mm / 2.01 / 1.89 / 0.63 m/s) 14,277 2,863 7,141 33.8 58.6 35.8 0.19
  • Steady RANS (FVM) retains stationary recirculation wakes behind pins and overpredicts wall temperature, whereas unsteady 3D thermal LBM resolves Kármán vortex-shedding mixing—lowering predicted temperatures closer to empirical correlations while quantifying unsteady hotspot fluctuations \(T_{\text{rms}}\) (0.19–1.92 K).