GPU AccelerationHPCParallel ComputingDifferentiable Simulation

High-Performance Computing & GPU Acceleration

A collection on making physics simulation fast: GPU-accelerated solvers and differentiable kernels, parallel computing and large-scale solves, memory and compute efficiency, and the underlying toolchain.

Updated Nov 3, 2026 9 resources
Introduction

Why this collection

Among the reasons physics AI projects fail, one class is not about the model — it is "we cannot run it": data generation too slow, training out of GPU memory, inference latency unacceptable. Only engineering addresses this class.

  • GPU acceleration — moving CPU code onto a GPU is not sufficient. GPU-ification of strongly coupled problems like algebraic multigrid or implicit solvers has its own methodology.
  • Parallel computing — partitioning and communication strategy for large sparse systems often determines total runtime more than single-node compute does.
  • Memory and compute efficiency — at scale, physics problems are often bound by memory bandwidth rather than floating-point throughput.
  • Toolchain — differentiable simulation kernels and GPU-native tensor libraries form the substrate on which any upper-layer method lands.

This collection is small, but every entry is genuinely usable in engineering: from mature solvers such as AMGX, to a custom implicit solver for full-3D tokamak electromagnetism, to memory-efficiency work on neural operators.

Start with the "GPU-Accelerated Frameworks & Solvers" group to build a mental map, then branch by your own problem type — strongly coupled, large sparse, or memory-bound.

GPU-Accelerated Frameworks & Solvers 3 resources

💬 Editor note: TokaGLINT is a custom implicit solver for full-3D tokamak electromagnetism — a deep case study on GPU-ifying a strongly coupled problem.

Architecture Acceleration of Machine Learning Gravitational Waveform Surrogate Models

Geophysics
This paper presents differentiable and hardware-accelerated implementations of the machine-learning surrogate models mlgw and mlgw-bns for gravitational-wave waveform modeling, enabling JIT compilation, vector parallelization, GPU acceleration, and automatic differentiation within JAX.
See LICENSE file gravitational waveswaveform surrogate modelsGPU acceleration
💬 Editor note: Inference acceleration at the architectural level is the route to putting a trained surrogate into a real-time loop.

Parallel Computing & Large-Scale Solves 3 resources

PETSc

Numerical Solver Library / HPC
High-performance scientific library from Argonne for large linear/nonlinear systems and parallel infrastructure - the underlying solver engine of top-tier CFD and multiphysics simulation, BSD-2-Clause friendly for commercial use.
New BSD-2-Clause SolverLinear AlgebraSparse Matrix
💬 Editor note: PETSc provides a GPU/CPU hybrid parallel linear algebra stack — the infrastructure for large sparse problems.

AMGX GPU-Accelerated Algebraic Multigrid Solver

HPC / Solver
NVIDIA's open-source distributed algebraic multigrid library offers classical Ruge-Stuben and unsmoothed aggregation AMG, multiple smoothers and Krylov methods, with a reported two- to fivefold single-GPU speedup over a competitive CPU implementation.
New BSD 3-Clause License (NVIDIA Corporation, 2017-2024) AMGXAlgebraic MultigridGPU Acceleration
💬 Editor note: AMGX is NVIDIA GPU algebraic multigrid and one of the de facto standards for accelerating sparse linear systems — worth using as a baseline.

A Parallel Solver for Vortex-Induced Vibrations at Zero Mass Ratio

CFD
The paper presents a parallel monolithic finite element strategy for simulating two- and three-dimensional vortex-induced vibrations of a rigid, massless cylinder mounted on translational springs.
See LICENSE file vortex-induced vibrationscylinder flowparallel solver

Memory & Compute Efficiency 2 resources

Adaptive Physics Transformer with Fused Global-Local Attention for Subsurface Energy Systems

Original Research Paper
ICML 2026 paper proposing APT, a geometry-/mesh-/physics-agnostic neural operator that fuses graph-based local feature extraction with global attention for long-range physical effects; achieves 4× saturation error reduction on 2D faulted CO2 storage vs prior SOTA, 19,873× speedup over GEOS and 6,800× over ECLIPSE.
arXiv preprint (copyright held by authors) Neural OperatorSubsurface EnergyCO2 Storage
💬 Editor note: APT adaptive attention reduces memory footprint while preserving long-range dependencies — a model of trading accuracy against overhead.

MENO: Memory-Efficient Neural Operator

ScientificML
A high-performance PDE neural solver based on Manifold Function Encoder, featuring resolution-independent memory footprint and arbitrary PDE input support for large-scale physical simulations.
See LICENSE file Neural OperatorPDE SolverMemory Efficiency
💬 Editor note: MENO attacks the memory bottleneck of neural operators directly — at large grid sizes this usually hits before compute does.

Toolchain 1 resources

NVIDIA Warp

Differentiable GPU Simulation Framework
NVIDIA's open-source Python framework that JIT-compiles plain Python functions into CPU/CUDA kernels with automatic differentiable adjoints, letting gradients flow through entire physics simulations - the compute substrate under Isaac Lab and Newton.
New Apache License 2.0 Differentiable SimulationGPUJIT Compilation
💬 Editor note: NVIDIA Warp offers a GPU-native differentiable simulation kernel and is the mainstream substrate for hybrid physics-ML simulation.