DevitoPRO is an enterprise version of the open-source Devito domain-specific language (DSL) that generates highly optimized computational kernels for seismic imaging and inversion workloads. It enables geophysics teams to write physics once and automatically regenerate tuned code for CPUs and GPUs across AMD, Arm, Intel, and NVIDIA architectures, as well as major cloud platforms. The platform includes advanced autotuning, multi-GPU MPI support, and large-model memory management for production HPC environments.
Funding
Funding not disclosed
Founders
Product
Problem
Seismic imaging and inversion workloads such as reverse time migration (RTM) and full waveform inversion (FWI) require computationally intensive PDE solvers that must be manually optimized for each new processor architecture. As hardware generations evolve rapidly across CPUs and GPUs, geophysics teams face significant engineering effort to maintain performance portability, often resulting in suboptimal kernels and delayed deployment on new infrastructure.
Solution
DevitoPRO provides an enterprise-grade symbolic DSL that lets users express PDE solvers once and automatically regenerate optimized kernels for any target architecture. The platform extends open-source Devito with production-focused tuning, portability, and support, generating CUDA, HIP, and SYCL code from a single symbolic specification. It includes advanced autotuning via DevitoTuner, GPU-aware MPI for multi-node scaling, and intelligent memory management with compression-based back-propagation and data streaming. DevitoPRO also offers benchmark reports, private support channels, and statement-of-work engagements to ensure reproducible performance across on-premises and cloud environments.
Target Audience
Primary customers are geophysics teams writing their own RTM/FWI kernels, seismic software vendors, and hardware designers who need production-grade, portable performance for seismic imaging and inversion workflows.
Features
- Symbolic DSL for expressing custom physics and PDE solvers with automatic code generation
- Advanced autotuning via DevitoTuner for speed-of-light performance on any architecture
- Multi-GPU and MPI support with GPU-aware communication, NUMA-aware MPI+OpenMP, and a decoupler for calling from any language
- Large-model memory management with compression-based back-propagation, disk-to-host-to-GPU streaming, and lossy floating-point compression
- CUDA, HIP, and SYCL backends generated from one symbolic spec for NVIDIA, AMD, and Intel GPUs
- Physics coverage including acoustic VTI/TTI, viscoacoustic VTI/TTI, elastic VTI/TTI, viscoelastic, and immersed boundaries for land topography
- Cloud deployment tuned and benchmarked across AWS, Azure, and GCP