High-performance and high-scalability have always been among the priorities when we develop our MCX toolset. The recommended hardware to get most out of MCX/MCXCL is to use discrete consumer graphics cards targeted for enthusiastic gaming users (medium to high-end), for example, GTX 980Ti, 1080, 1080Ti for NVIDIA, Radeon RX 480, Nano, 580, Vega 56/64 for AMD cards. These cards typically cost between $200 and $500. MCX/MCXCL have been extensively optimized on these cards to maximize the single-precision floating-point (fp32) computation throughput.
We DO NOT recommend purchasing and using the more expensive Tesla GPUs (P100, V100) for MCX simulations. This is because MCX/MCX-CL were explicitly optimized not to use double-precision (fp64) computation, which is the main strengths of these lines of products. In our past benchmarks, Tesla cards generally performs comparably or slightly lower in speed compared to the highest end consumer GPUs of the same generation. Of course, if you already have purchased those GPUs, MCX/MCXCL will perform well.
| || |
MCX Speed Benchmark (for NVIDIA GPUs)
| || |
MCX-CL Speed Benchmark (for NVIDIA/AMD/Intel processors)