The HPC Trilemma: Choosing Between Performance, Portability, and Productivity
The physical sciences describe the world with applied mathematics. Seismology, electromagnetics, and fluid dynamics all rely on techniques like partial differential equations and finite-difference simulations to model physical phenomena.
Python and its scientific ecosystem of libraries like NumPy, SciPy, and MatPlotLib, give scientists and researchers frameworks to develop sophisticated domain-specific solutions using relatively few lines of readily accessible code. It is this versatility and productivity that has propelled Python into one of the leading languages in science and engineering today. In what other programming language can you literally import antigravity?
While pure Python may be slower than compiled languages, its primary role in high-performance computing (HPC) is as an accessible interface to highly optimized libraries written in lower-level, compiled languages such as C, C++, Fortran, and even CUDA, SYCL, or HIP for GPU support. Python’s scientific libraries, like NumPy and SciPy, leverage these compiled backends to deliver high performance while maintaining Python’s ease of use. However, this model still depends on the availability of optimized HPC code for domain-specific tasks, which often requires specialized expertise and resources to develop and maintain.
The Devito Solution: Let Scientists Be Scientists
The Devito Codes team aimed to bring just-in-time (JIT) compilation of HPC-grade, finite-difference code to the Python ecosystem. With Devito, scientists can work within Python’s symbolic and mathematical framework (SymPy), writing complex partial differential equation solvers and goal-driven optimization problems, and seamlessly generate parallelized, hardware-optimized HPC code for all major CPU and GPU architectures.
Devito Codes See Major Gains with Intel Xeon 6 (higher is better)
A Devito Codes kernel for an acoustic anisotropic propagator for tilted transverse isotropy (TTI) model (a typical energy industry seismic model) illustrates Intel Xeon 6 performance gains over 5th Gen Intel Xeon Scalable processors.1
“Most people adopt Devito initially for the productivity boost,” says Gerard Gorman, CEO and Co-founder of Devito Codes. “It enables a unique form of rapid prototyping, letting them treat it as a computational laboratory.”
Creating a Python framework that can produce high-performance kernels for simulations, inversion, and optimization tasks across diverse hardware isn’t straightforward. Devito integrates multiple HPC technologies, including OpenMP for shared memory systems, OpenACC for accelerators, and MPI for parallelism and portability. Advanced optimizations require hardware-specific tuning with specialized languages like CUDA, HIP, and SYCL. Devito Codes applies nearly every known optimization technique for structured computation and continuously integrates new advancements, ensuring that performance gains compound over time—often rivaling or even surpassing expert hand-tuned commercial solutions.
By taking on the complexity of HPC code development, Devito Codes enables users to transfer projects across systems effortlessly, leveraging all available computing resources. HPC operators and service providers can maintain current infrastructure, expand with heterogeneous arrays, and upgrade significantly—all while ensuring that workloads remain compatible and performant.
“With DevitoPRO, a geophysicist can take an algorithm from a research paper and implement it in an afternoon —a task that would normally take months of coding and optimization. This rapid turnaround allows teams to experiment and innovate at an unprecedented pace, accelerating the testing and deployment of new algorithms.” —Mathias Louboutin, Senior Solutions Architect, Devito
DevitoPRO
Devito began as part of an Intel Parallel Computing Centre initiative led by Professor Gerard Gorman at Imperial College London. The initial project created high-performance, open-source software for seismic imaging. As the project matured into a true optimizing compiler for HPC workloads, the team launched DevitoPRO, an enterprise edition with proprietary features, advanced performance optimizations, and commercial support.
DevitoPRO primarily serves exploration geophysics in the energy industry. In addition to compiling highly optimized, portable code for seismic simulations authored in Python, DevitoPRO provides high-performance propagators and gradient operators for Full Waveform Inversion (FWI) and Reverse Time Migration (RTM). DevitoPRO also provides technical support, training, custom software development, and hardware-specific optimization for clients.
Devito continues to provide general-purpose symbolic and compiler software technology as open-source, patent-free software for researchers in academia and industry.
Expanding Code Portability with SYCL and Intel
Traditionally, creating portable code that can run across heterogeneous processors required compiling unique kernels for each hardware type—CUDA kernels for NVIDIA GPUs, HIP kernels for AMD GPUs, and C/C++ for x86 and RISC CPUs. In recent years, SYCL has given HPC programmers a new, multi-platform option for compiling optimized HPC code.
“Devito’s fusion of symbolic computation and advanced compiler technology ensures reliable, verifiable, optimized code generation—critical for mission-focused mathematical software. While generative AI code lacks this level of precision and dependability, combining both technologies can unlock even greater productivity in developing and testing new algorithms.” —Gerard Gorman, CEO and Co-founder of Devito Codes
SYCL is a cross-platform parallel C++ abstraction layer with APIs that can find and manage data resources and code execution on mixed devices from multiple vendors including CPUs, GPUs, and FPGAs. SYCL is the foundation for moneAPI and Data Parallel C++, which Intel has implemented on Intel® Data Center GPUs.
Devito Codes and Intel engineers worked together to bring SYCL code generation to DevitoPRO, including specific optimizations for Intel® Data Center GPU Max 1100 and 1550 series accelerators. To deploy, DevitoPRO users simply target Intel Data Center GPUs for just-in-time compilations that reap the performance benefits of SYCL.
Expanding Performance for Elastic Wave Seismology with Mixed Precision Computing
Seismic waves travel in two forms: longitudinal primary waves (P-waves) and transverse secondary waves (S-waves). Modeling P-waves—a method seismologists call acoustic analysis—is relatively simple mathematically and computationally because the wave energy and particle motion travel in the same dimension. Modeling P- and S-waves together—which seismologists call elastic analysis—complicates things geometrically.
Because the two waveforms move in three dimensions at ninety-degree angles, describing them requires more wave equations with more components. Elastic analysis also requires much higher resolution to produce accurate results, which means gathering vastly more data.
Seismic waves come in two forms: longitudinal Primary (P) waves and transverse Secondary (S) waves.
“If you are running elastic, then your memory footprint is going to be, in the best case scenario, between four and five times larger than in the case of acoustic,” says Fabio Luporini, CTO and Co-founder of Devito Codes. “It’s simply because of the physics. You are modeling more wave fields simultaneously in a coupled partial differential equation, which you have to keep in memory.”
Devito Codes is developing mixed-precision algorithms, an AI computing technique, to make elastic computing workloads possible on current-generation hardware. Workloads that can afford a small loss in precision are converted from FP-32 (32-bit floating point) to a carefully designed mix of FP-32 and FP-16 (16-bit floating point), which represent the same values with half the memory. In the world of elastic analysis, halving a petabyte dataset to 500TB produces cascading performance boosts at every step from memory and I/O management to writing snapshots to disk.
Devito Codes Doubles Performance with Mixed Precision (higher is better)2
Devito Codes tests show shifting and mixed precision (FP-16/FP-32) increases performance 2x and reduces memory footprints 2x, resulting in significantly faster throughput.2
FP-16 workloads also process faster on CPUs and GPUs that support mixed precision, like Intel® Xeon® 6 processors and Intel Data Center GPUs. In initial tests, elastic wave analysis running in mixed precision with Devito Codes produced a up to 2x performance increase,1 the equivalent of a step change in performance with no hardware upgrades.
Conclusion: Every Performance Increment Counts
Devito Codes and Intel continue to refine and optimize compiler technologies to extract more performance from heterogeneous HPC systems for finite difference simulations. For Devito’s open-source and DevitoPRO clients, the work is indispensable.
“Processing seismic data for subsurface imaging can cost millions in compute expenses per project,” says Gorman. “So we need to squeeze out every last percent of performance because time is money.”
For the latest updates, visit Devito Codes on GitHub or devitocodes.com.
