Writes are atomic in redis because redis is single threaded. So you are bounded by how fast redis can write. If you try to write any faster then redis can handle you'll get queueing or errors.
The wonderful thing about HyperLogLogs is that you can split the counter in N servers and "merge" the registers later, in case you want an architecture that shards the same counter in multiple servers. But sharding directly by resource looks simpler actually...