Why are GPUs so much faster for some physics codes but not others?

GPUs excel when the same simple operation is applied to huge amounts of data with regular memory access, as in force or stencil updates. Codes with heavy branching, irregular memory access or large serial sections see little benefit and can even run slower.

Do GPUs replace CPUs in physics computing?

No. Most large simulations are heterogeneous, using GPUs for the heavy data-parallel kernels while CPUs handle control flow, coordination and irregular work, so the two are used together rather than one replacing the other.

GPU and Accelerator Computing in Physics

Graphics processing units and other accelerators pack thousands of lightweight cores that excel at the data-parallel arithmetic of physics simulation, delivering large speedups for molecular dynamics, lattice and grid computations.

Znajdź temat z PaperMindWkrótceFind papers & topics

Tools & resources

Pobierz slajdy

Learn & explore

WideoWkrótce

Definition

GPU and accelerator computing in physics is the use of highly parallel coprocessors with many simple cores to execute the data-parallel kernels of a simulation, achieving high throughput on suitable physics workloads.

Scope

This topic covers accelerator computing for physics: the massively data-parallel GPU model, memory hierarchy and the importance of memory-access patterns, mapping physics kernels such as force computation and stencil updates onto accelerators, and the trade-offs of heterogeneous CPU-GPU programming. It addresses both performance and programmability.

Core questions

Why do GPUs suit the data-parallel arithmetic of many physics simulations?
How does the GPU memory hierarchy shape achievable performance?
Which physics kernels map well onto accelerators and which do not?
How are heterogeneous CPU-GPU simulations organized?

Key theories

Massively data-parallel execution: GPUs run the same operation across thousands of threads on different data, matching simulations where identical updates are applied to many particles or grid points, which yields large throughput gains.
Memory hierarchy and access patterns: GPU performance hinges on coalesced memory access and effective use of fast on-chip memory, so algorithms must be restructured to keep the many cores fed with data rather than stalling on memory.
Accelerated physics kernels: Force evaluation in molecular dynamics, stencil updates in grid solvers and lattice updates in field theory have been ported to GPUs for order-of-magnitude speedups, as demonstrated by early fully GPU-based molecular dynamics.

Clinical relevance

GPU acceleration powers much of modern molecular dynamics, lattice quantum chromodynamics and astrophysical simulation, and the same hardware underlies machine-learning workflows increasingly used to analyze and accelerate physics computations.

History

General-purpose GPU computing took off in the late 2000s with programmable shaders and the CUDA platform; physics was an early adopter, with fully GPU-based molecular dynamics in 2008 and GPU-accelerated lattice quantum chromodynamics following, making accelerators standard in supercomputers.

Key figures

Wen-mei Hwu
David Kirk
Joshua Anderson

Seminal works

kirkhwu2016
anderson2008

Frequently asked questions

Why are GPUs so much faster for some physics codes but not others?: GPUs excel when the same simple operation is applied to huge amounts of data with regular memory access, as in force or stencil updates. Codes with heavy branching, irregular memory access or large serial sections see little benefit and can even run slower.
Do GPUs replace CPUs in physics computing?: No. Most large simulations are heterogeneous, using GPUs for the heavy data-parallel kernels while CPUs handle control flow, coordination and irregular work, so the two are used together rather than one replacing the other.