The benefits of speculative decoding depend on the system having surplus compute capacity; NVIDIA GPUs, whose compute far exceeds memory bandwidth, are naturally suited to this optimization, while Apple chips' more balanced compute-to-bandwidth ratio results in lower gains.