Partial differential equations (PDEs) underlie much of the natural world and are ubiquitous in engineering and scientific modeling. Obtaining accurate and precise solutions to PDEs is an important and notoriously difficult / expensive problem.
To build the PDEs into the loss function, partial derivatives of the prediction with respect to the input data are calculated by automatic differentiation. Calculated in this way, the depth of the computation graph grows exponentially in the order of the PDEs.
These deep computation graphs are well-suited to dataflow architecture. Instead of proceeding kernel by kernel through the computation graph (slow), SambaNova Reconfigurable Dataflow Units™ (RDUs) allow for data to be pipelined, enabling high compute utilization / faster training / etc.