[mikechen233]

It's not due to the on package or the united memory. M1 is using industry standard 128bits lpddr4. On package is still using regular dram chips. What seemingly is the advantage is the firestorm cpu has much wider pipeline and capable doing more load and store in-flight at the time. Also the cache is also able to provide low latency and the bandwidth numbers. Intel or amd is able to achieve similar performance in memory bound workloads if they designed the logic on cpu for that workload.