Poor performance when comparing bulk loading to SQL Server on-premise versus SQL Server APS/ADW (Analytics Platform System/Azure Data Warehouse).

It is not valid to compare insert performance to ADW, Azure SQL database, and SQL Server on-premises.

ADW performance is constrained by the fact that bulk load to all distributions (ex. 60) will go through the control node. 

Depending on the DWU (Data Warehouse Units) setting, there will be between 1 and 60 compute nodes with 60 to one database distributions connected to each node. The fact that the data is being written to 60 different files alone will change the performance profile. With more than one compute nodes, the DMS service (Database Migration Service) needs to move the data from the control node to the compute node(s) to persist the data, which introduces additional overhead. 

That said, bcp performance and bulk insert performance into the same platform and table geometry should yield similar results. Smaller row counts would increase the chatter on the wire and would likely slow down the overall transfer speed. 

When writing into a table with a CCI (Clustered ColumnStore Index) – ideal row group quality is achieved with 1 million records per distribution – the compression of the row group obviously causes some overhead on the compute node(s) and can affect load performance. The larger impact of poor row group quality would not show up until you actually start retrieving information from the tables. 

There is a way of parallel loading from multiple text files on individual compute nodes, but this mechanism is only available through the Polybase external table functionality.