Core pattern
As seismic algorithms moved from stacking to migration, RTM, and FWI, compute requirements increased by orders of magnitude.
Data growth
Typical seismic projects grew from analog records and megabyte-scale digital surveys to modern 100 TB to multi-petabyte programs.
Architecture shift
Mainframes gave way to vector systems, then distributed-memory clusters, and finally GPU-dense platforms with very high I/O throughput.
Why oil and gas mattered
Seismic processing consistently pushed memory bandwidth, storage bandwidth, and large-scale numerical simulation harder than most commercial workloads.
How the algorithms evolved
The cleanest way to understand this history is to track the progression of seismic algorithms. As computing power increased, the industry moved toward more complete representations of wave physics.
| Stage | What it added | Relative cost |
|---|---|---|
| Migration | Transforms recorded events into a spatial image of the subsurface. | 1× baseline |
| Reverse Time Migration | Uses the full two-way wave equation and handles complex geology much better than simpler migration methods. | ~10× |
| Acoustic FWI | Fits full wavefields to update velocity models iteratively. | ~100× |
| Elastic FWI | Adds shear physics and more complete subsurface behavior. | ~500–1000× |
| Multi-physics inversion | Pushes toward richer coupled models and even larger optimization problems. | 1000×+ |
Processing Workload Advances
Seismic HPC repeatedly evolved across the same planes: compute, network, storage, control, and facility power. Algorithm demands shaped each of them.
| Era | Channels | Traces per Survey | Workload | Processing Scale |
|---|---|---|---|---|
| 1930s | 6-12 | ~1000 | Manual interpretation (travel-time measurement, hand calculations) | Concentrating really hard |
| 1940s | 12-24 | ~5,000 | Manual plus mechanical calculators | Discussing with others |
| 1950s | 24-28 | ~20,000 | Early analog filtering and stacking experiments | Analog Kiloscale |
| 1960s | 48-96 | ~100,000 | Digital stacking and velocity analysis | Megascale |
| 1970s | 96-240 | ~1 million | Digital filtering, deconvolution, NMO stacking | Tens of Megascale |
| 1980s | 240-1000 | ~10 million | Early 3D processing and migration | Gigascale |
| 1990s | 1000-3000 | ~100 million | Large-scale 3D migration | Tens to Hundreds Gigascale |
| 2000s | 3000-10,000 | ~1 billion | Prestack time migration | Terascale |
| 2010s | 10,000-50,000 | ~10 billion | Reverse Time Migration | Petascale |
| 2020s | 50,000-200,000+ | ~100 billion | RTM+FWI and AI-assisted | Exascale |
Data growth Forcing Architectural Changes
The rise of HPC in oil and gas was not driven by compute alone. Acquisition systems captured more channels, denser sampling, longer records, and larger areas. That created a storage and network problem just as serious as the CPU or GPU problem.
| Era | Typical data size | Dominant acquisition pattern | Processing implication |
|---|---|---|---|
| 1950s | KB to MB equivalent | Analog 2D reflection surveys | Manual and analog processing |
| 1960s | MB to 100 MB | Digitized 2D surveys on magnetic tape | Mainframe batch processing |
| 1970s | 100 MB to 1 GB | Multi-channel marine acquisition | Industrial tape-library workflows |
| 1980s | 10 GB to 100 GB | Commercial 3D seismic surveys | Vector supercomputers for large numerical methods |
| 1990s | 100 GB to 1 TB | Larger 3D marine surveys | Massively parallel computing and distributed storage |
| 2000s | 1 TB to 10 TB | Wide-azimuth and ocean-bottom systems | Linux clusters and parallel filesystems |
| 2010s | 10 TB to 100 TB | High-density nodal and monitoring deployments | GPU acceleration and high-throughput storage |
| 2020s | 100 TB to multi-PB | DAS, permanent monitoring, ultra-dense arrays | Exascale-class hybrid HPC systems |