It isn't obvious to me that they intended to give this as the maximum single-core performance, or just the proportional share of 844GB/s across 136 cores. Implementations of Neoverse V2 by Nvidia and Amazon hit 20-30GB/s in single-threaded work.