While asymptotic solvers like Savant have smaller computational requirements than full-wave techniques, the costs increase significantly with frequency and simulation fidelity and benefit greatly from acceleration techniques.
Graphics processing units (GPUs) are throughput oriented processing devices that are well-suited for the mathematically intensive workloads found in CEM solvers. GPUs contain hundreds of processing units and leverage thousands of threads.
The GPU acceleration in Savant is a hybrid CPU and GPU approach that executes on all CPU cores and GPUs present in a machine. It can also be extended to execute on multiple nodes in GPU-based clusters using MPI. Current results for the GPU parallelization of Savant show acceleration factors of up to 217 times using 1 GPU card and a quad-core CPU and up to 380 times using 2 GPU cards and a quad-core CPU. Larger acceleration factors than those quoted here can be realized utilizing more GPU cards and CPU cores.
Large Savant computations that use to take hours or even days to run now complete in seconds or minutes with no compromise in accuracy. This massive acceleration of the Savant computational engine helps users to more effectively explore trade spaces by rapidly generating high fidelity simulation data for many different scenarios.
The Savant GPU feature requires NVIDIA GTX, Quadro or Tesla GPU cards.