Beyond the GPU: Rethinking Seismic Compute with Dataflow

Elad Raz, Founder & CEO, NextSilicon

The global energy transition is well underway. Renewable sources are expanding, electrification is accelerating, and investment in sustainable infrastructure is reaching new heights. Yet the reality is clear. Oil and natural gas will remain essential to the global economy for decades to come. That makes efficient, responsible resource extraction more important than ever. At the heart of this effort lies high-performance computing (HPC) and the seismic simulations that guide exploration, production, and refinement decisions with multi-billion-dollar implications.

The high-performance systems running seismic simulations such as reverse time migration (RTM), full-waveform inversion (FWI), and other advanced modeling processes must evolve to extract smarter. That means reducing environmental impact, maximizing reservoir recovery, minimizing operational risk, and making every barrel count. 

But the computational infrastructure supporting these workloads is reaching its limits, not because we lack processing power, but because we’re using the wrong architecture for the job. 

The Seismic Cost of Seismic Compute

Seismic imaging is among the most computationally intensive applications in scientific computing. FWI can be hundreds to thousands of times more computationally expensive than RTM. As exploration moves to deeper water, more complex geology, and unconventional reservoirs, the need for high-fidelity models drives these computational requirements to new heights.

For the past decade, the industry has turned to GPU acceleration. And for good reason. GPUs have enabled dramatically higher floating-point throughput at attractive performance-per-dollar ratios. Today, the largest seismic processing centers run on thousands of GPUs, processing petabytes of survey data to generate the subsurface images that drive exploration decisions.

The scale of this challenge is staggering. A single high-resolution 3D FWI run on a leading supercomputer required roughly 25,000 GPU-hours across hundreds of GPUs (enough energy to power one average U.S. home for 15 months). For operators running these workloads in private data centers or the cloud, that translates directly into operational cost, power consumption, and infrastructure complexity.

Yet despite massive GPU deployments, many operators report that seismic workloads achieve only 50% of theoretical GPU performance. Expensive accelerators sit partially idle while energy consumption and cooling demands continue to climb.

The problem isn’t a lack of compute. It’s an architectural mismatch.

Why GPUs Struggle with Seismic Compute

Over the past decade, GPUs have become the accelerator of choice for HPC. Their high floating-point throughput and strong performance-per-dollar made them attractive for scientific workloads.

But seismic imaging isn’t primarily compute-bound. It’s memory-bound.

Seismic wave propagation algorithms, which calculate values at each grid point based on neighboring points, repeatedly move large volumes of data between memory and processing units. Every set of calculations requires fetching data from neighboring grid points, processing it, and then repeating this cycle billions of times. The arithmetic itself is relatively simple; the bottleneck is data movement.

GPU manufacturers optimize their architectures for workloads where data, once loaded, can be reused heavily for computation. In memory-bound seismic workloads, that reuse is limited. The result: powerful compute units spend much of their time waiting for data to arrive from memory.

At the same time, new GPU generations increasingly prioritize AI workloads, where lower-precision operations dominate. Seismic modeling, by contrast, relies heavily on high-precision FP64 performance. This creates what many operators describe as an “AI tax”, paying for silicon and power optimized for workloads that don’t align with seismic needs.

The industry has responded by scaling out: more GPUs, larger clusters, higher power draw. But simply building bigger GPUs or faster memory doesn't solve this fundamental problem; it just moves the bottleneck.