[VHRanger]

Yeah when benchmarking by batch sizes it's common to see huge jumps associated with the memory hierarchy:

- word size (64bits) - cache alingment fetch size (generally 64bytes as mentioned above) - OS page size (4-16kb) - L1 size (~80kb/core) - L2 (low megabyte number)

[hinkley]

With lots of bizarre artifacts if you don’t force alignment.