GPU Acceleration in Maxwell3D

• Graphics Processing Units (GPUs) are supported by the complex matrix solver in the Maxwell 3D eddy current solver. GPU support is also available for the real matrix solvers in the Maxwell 3D magnetic transient, magnetostatic, DC conduction, electrical transient, and electrostatic solution types.
• Performance depends upon matrix size as well as hardware. When the matrix size is large (usually more than 2 million DOF), you should see a performance improvement if a GPU is used.
• Example speedup factors from Ansys AEDT simulations showed 1.25X for 2.8M unknowns, 2X for 5.3M unknowns, and 2.5X for 6.6M unknowns. Another example model realized 1.3X speedup for a P100 GPU used with 8 cores for an unspecified matrix size.
• GPU support relies on the offload model so a CPU is still required. Certain key computations in the multifrontal sparse direct solvers are offloaded to the GPU for acceleration. The solvers only use GPUs when there can be a potential speedup.
• The Maxwell3D Eddy Current solver can also use multiple distributed GPUs for solving frequency sweeps.
• Ansys HPC licensing is required to use a GPU. For Ansys electronics products, 1 GPU is unlocked by every 8 HPC tasks. 1 HPC Pack enables 12 CPU cores + 1 GPU, 2 HPC Packs enable 36 CPU cores + 4 GPUs, etc.