The memory bandwidth bottleneck limits the speed of running local models, the fact that this model is parallelizable means that even with one batch inference it

The memory bandwidth bottleneck limits the speed of running local models, the fact that this model is parallelizable means that even with one batch inference it will be possible to balance memory bandwidth bottleneck and compute bottleneck (aka much more speed).