I think SHA-256 is still marginal for speed in modern environments unless your I/O is unusually limited relative to CPU. Current servers can support 10s of GB/s combined throughput for network and storage, which is achievable in practice for quite a few workloads. Consequently, you have to plan for the CPU overhead of the crypto at the same GB/s throughput since it is usually applied at the I/O boundaries. The fact that SHA256 requires burning the equivalent of several more cores relative to Blake3 has been a driver in Blake3 anecdotally creeping into a lot of data infrastructure code lately. At these data rates, the differences in performance of the hash functions is not a trivial cost in the cases where you would use a hash function (instead of e.g. authenticated encryption).
The arm64 server case is less of a concern for other reasons. Those cores are significantly weaker than amd64 cores, and therefore tend to not be used for data-intensive processing regardless. This allows you to overfit for AVX-512 or possibly use SHA256 on arm64 builds depending on the app.
There is a strong appetite for as much hashing performance per core as possible for data-intensive processing because it consumes a significant percentage of the total CPU time in many cases. Due to the rapid growing scale, non-cryptographic hash functions are no longer fit for purpose much of the time.
The arm64 server case is less of a concern for other reasons. Those cores are significantly weaker than amd64 cores, and therefore tend to not be used for data-intensive processing regardless. This allows you to overfit for AVX-512 or possibly use SHA256 on arm64 builds depending on the app.
There is a strong appetite for as much hashing performance per core as possible for data-intensive processing because it consumes a significant percentage of the total CPU time in many cases. Due to the rapid growing scale, non-cryptographic hash functions are no longer fit for purpose much of the time.