We used tables because we have 2 types of queries on this table:
SELECT * FROM table_wih_lots_of_columns WHERE condition involving a lot of columns.
These type of queries read lot of rows.
or
SELECT columnX FROM table_wih_lots_of_columns WHERE condition involving a lot of columns
These type of queries read very few rows.
Everyone else has had great advice on this, I'd like to add that arrays of any dimension are limited to 1GB like all varlena objects.
You should check out pg-strom, it's highly optimized for running exactly these kinds of queries on a GPU and comes with a native matrix type that can exceed the 1GB limit.