It was probably just before running LLMs with tensor parallelism became interesting. There are plenty of other workloads that can be divided by 6 nicely, it's not an end-all thing.