Computational Throughput of Accelerator Units with Application to Neural Networks

The size of data that can be fitted with a statistical model becomes restrictive when accounting for hidden dynamical effects, but approximations can be computed using loosely coupled computations mainly limited by computational throughput. This whitepaper describes scalability results attained by implementing one approximate approach using accelerator technology identified in the PRACE deliverable D7.2.1 [1], with the aim of adapting the technique to future exascale platforms