I will agree with him that the unified memory architecture is more efficient, however the amount of efficiency is grossly over exagerated.
The issue with the UMA to "traditional" memory comparison (and to a larger extent the M1 to x86 comparison as a whole) is that there is no real way to quantify how much of a difference UMA makes compared to a traditional memory setup. It's not just that the RAM is on the SoC instead of on the other end of the system bus, but also the fact that the RAM is not partitioned. On any x86-based system with an iGPU, part of that system RAM is partitioned off for graphics. So if there is data that both the CPU and GPU have to work on, it is copied twice (once per partition), then the data is manipulated as needed and passed back through the system bus (at which point the two data sets are compiled into a unified set). It actually is at a minimum 2x the work of a unified architecture. With UMA, the data is copied to the system RAM once, and both the GPU and CPU can access that simultaneously. This speeds up the process in two ways: the system no longer has to pass data back and forth through the system bus, and since the CPU and GPU can handle the data simultaneously, there is no need to compile two versions of the same data into one unit.
The other factor that plays into the overall performance of the M1 is the speed of the SSDs Apple is now using. They are fast enough (near-DDR4 speeds) that using the swap file has a very minute impact on performance. In a way it's ironic that at the same time Intel is boosting core clock speeds of their CPUs to create the illusion of improved performance, Apple is boosting data transfer speeds on both the RAM and storage sides of the equation and achieving comparable (if not superior) performance to x86 systems at a lower clock speed, while also using significantly less battery in the process.
While the M1's RAM setup is definitely more efficient than the traditional model, it's hard to quantify that efficiency in terms of actual performance. For some apps, 8GB in the M1 may work similarly to 12 GB in an x86 machine. For other apps, that 8GB in the M1 could work identically to 8GB in an x86 machine. This variable is really dependent on how the app was coded to handle both RAM and swap space.
Last edited: