I think you can definitely improve the throughput/iops by using BAR vs treating it like a file store/mount through cuda which adds a lot of overhead.