The Data Parallel Haskell work on flattening ran into problems with space blow-up due to replicating arrays to perform flattening. Does Futhank avoid those problems?
(That said, memory explosion is still a common symptom when the compiler misjudges how much parallelism should be exploited.)