Numbers Don't Lie
We benchmark every Permeon release against CPU reference runs. Here are the results — unfiltered. Methodology and hardware specifications included.
Runtime Comparison
CPU baseline: 2× Intel Xeon Gold 6348 (28 cores each, 56 total threads). OMP_NUM_THREADS=56. Reference simulator: OPM Flow v2024.10.
| Model | Cells | Timesteps | CPU Time | GPU Time | Speedup | GPU |
|---|---|---|---|---|---|---|
| SPE-1 | 300 | 1,200 | 18s | 0.3s | 60× | A100 |
| SPE-9 | 9,000 | 900 | 4 min | 3s | 80× | A100 |
| Synthetic Field A | 150,000 | 1,500 | 45 min | 30s | 90× | A100 |
| Synthetic Field B | 500,000 | 3,000 | 6 hr | 4 min | 90× | H100 |
| Large Field C | 2,000,000 | 5,000 | 48 hr | 28 min | 103× | H100 |
All GPU times are wall-clock including data transfer. CPU times use optimal thread count. Models are synthetic test cases; contact us for your field-specific benchmark.
GPU Scaling by Model Size
Typical runtime for a 10-year simulation with adaptive timestepping. Results vary by model complexity, fluid properties, and well count.
| Grid Cells | NVIDIA A100 (80GB) | NVIDIA H100 (80GB) |
|---|---|---|
| 10K | 1.2s | 0.8s |
| 100K | 12s | 7s |
| 500K | 4 min | 2.5 min |
| 1M | 12 min | 7 min |
| 2M | 48 min | 28 min |
| 5M | 3.5 hr | 1.8 hr |
Methodology
CPU Reference Setup
- ▸ 2× Intel Xeon Gold 6348, 28 cores each
- ▸ 512 GB DDR4-3200 ECC RAM
- ▸ OPM Flow v2024.10 (open-source Eclipse-compatible)
- ▸ OMP_NUM_THREADS=56, optimal DUNE AMG settings
GPU (Permeon) Setup
- ▸ NVIDIA A100 SXM4 80GB or H100 SXM5 80GB
- ▸ CUDA 12.4, cuSPARSE, custom CUDA kernels
- ▸ CPR preconditioner with cuAMG pressure stage
- ▸ Wall-clock time includes all data transfer overhead
Accuracy Criterion
All GPU runs are compared to the CPU reference for FOPT, FWPT, FGPT, and WBHP. Maximum relative deviation must be below 1.0% at any reported timestep. Runs failing this criterion are not reported.