I like this image of the Windows storage driver stack.
The allocation unit is an attribute of the filesystem, governing initial the incremental filegrowth possible.
Although its old, I like this explanation of how the NTFS filesystem works as a whole.
How an individual file is accommodated in the NTFS filesystem is explained very thoroughly in these two posts.
Once clusters are allocated to a file for growth, the work of the allocation unit is done. When a physical read is issued against a file, at the filesystem layer the file-based address is resolved for lower levels of the Windows stack based on the cluster allocations.
Once the file has grown, the remaining questions are how compact its metadata is and how contiguous its data is.
A single file on a new volume should have completely contiguous data. Unless the NTFS metadata was fragmented and bloated by a lot of prior activity like massive amounts of small file creation and deletion, the metadata should also be very compact.
With contiguous clusters and compact metadata, whether allocations were made to the file in 4k clusters or 64k clusters can end up falling out of the equation.
If that file is a transaction log, sequential writes of 512bytes to 60k will occur regardless of 4k/64k cluster size. If its a database file, reads of 8k to 512k (the default Windows maximum physical IO size and maximum physical IO size for VMware pvscsi vhba) will also occur regardless of 4k/64k cluster size. On a VMware vm, vmfs filesystem block size will have minimal effect outside the first write to thin volumes. The 1 mb vmfs block size doesn't force small physical reads to become larger.
On real systems there are often departures from the ideal. Multiple files within a filesystem increase the odds of fragmentation of both data and metadata. Sometimes that's not even intentional. If a SQL Server data file grows while DBCC checkdb is running, the sparse file for the database snapshot can result in fragmentation of the base database file. Sparse files themselves are very susceptible to fragmentation, as is discussed in the "four stages part 2" post above.
In the past, NetApp indicated observed performance improvements on systems with 4k allocation units vs 64 k allocation units. I don't doubt their results, but I am skeptical of their conclusions. I bet closer examination of the test systems would reveal something other than cluster size was responsible for observed performance difference. Maybe file fragmentation within the filesystem. Maybe filesystem metadata fragmentation or bloat. Maybe 4k vs 512byte disk sector size...
4k allocation units allow large numbers of small files to be stored in an NTFS filesystem without a lot of empty space in the allocated clusters. Although there have been claims of observed performance gains for SQL Server with 4k au, I've always suspected there's another cause in those cases.
64k allocation units allow much larger maximum Windows volume sizes than 4k allocation units. For systems like SQL Server where small numbers of large files are expected within an NTFS filesystem, 64k au seems a natural choice. Its much less likely to experience very badly fragmented metadata and hopefully (if more than 1 file in the filesystem) will experience greater contiguity. But predicting a performance gain on any given system by converting from 4k to 64k au AND pinpointing the reason for improved performance is a very tricky business. I'm not overly skeptical when I see evidence of gain on some systems. But I'm also not surprised to see legitimate tests on other systems that show no measurable difference.