GPU Cluster solver is an extension of GPU Solver, which combines GPU and MPI parallelization, in order to speed-up simulations and enables solving of electrically larger problems, than it is possible on a single machine. Advanced parallelization algorithms which are implemented enable the solver to very efficiently utilize an arbitrary number of compute nodes and an arbitrary number of GPUs on each node.

GPU cluster is intended for fast solving of multi-frequency problems of medium or relatively large size, and solving of electrically extremely large problems. In that sense, there are two operation modes of WIPL-D GPU Cluster Solver:

  • Single-node solver – each cluster node simulates independent frequency samples or independent projects
    • Multi frequency project is automatically divided into single frequency projects, which are run on different cluster nodes
    • Acceleration proportional to number of nodes


  • Multi-node solver – multiple nodes are employed in solving of one problem
    • All simulation steps (matrix fill-in, matrix inversion and post-processing) are performed on all employed nodes
    • Simulation time is reduced nearly number of nodes times
    • Maximum dimension of the solvable problem is significantly increased


Since each cluster configuration is unique, our solution is custom-made to best fit both, the available hardware and customer’s budget. For more info on our GPU cluster solution, please contact our Tech support.

Below are shown some examples of projects that were simulated on custom-made GPU cluster using the appropriate solver.

Configuration of the cluster used for the simulations

Example 1 – Monostatic RCS of global hawk

  • Frequency of interest: 5 GHz
  • Maximal dimension of the aircraft: 36 m (600 λ)
  • Output results: monostatic RCS calculated in 1801 directions, in vertical plane
  • One symmetry plane is used
  • Number of unknowns: 546,000
  • Simulation time (cluster with 8 nodes): 8 hours






Example 2 – Microstrip patch antenna at helicopter fuselage

  • Frequency of interest: 2.7 GHz
  • Helicopter length: 19 m (171 λ)
  • Number of unknowns: ~404,253
  • Simulation time (cluster with 6 nodes, each node has 2 GPUs): 7.7 hours