Not that different. DM-SMR needs you to write about 256-512M sequentially before FUA or they'll journal the write which positively annihilates throughput once the journal is full. Erase blocks in a modern SSD are probably still somewhat smaller than that, but not by much (32-64M?). With DM-SMR it's also important to avoid interleaving writes. Writing parallel streams of sequential chunks tanks performance since it starts journaling. At least that's what it looked like when I was developing some code that had to deal with them some years ago.