The problem comes down to the use of QLC flash memory for storage over TLC. As the names imply, QLC can store four bits of data per memory cell, while TLC stores three bits. Every time a single bit changes, all four bits need to be rewritten with QLC SSD media.

With the creation of 3D NAND, TLC memory is considered a good choice if you want to balance the speed of accessing stored data with capacity.

There's also durability to consider. Due to writing more bits to each memory cell, QLC is both slower and less durable over time than TLC memory.

To get around these limitations, Apple included a small cache of single-level cell (SLC) SSD media, which handles brief write commands and acts as a buffer with the main storage capacity. This cache is extremely fast, in the region of 3,000MB/s, so long as it has available room in its very small capacity.

Once this is exhausted, there's a second buffer of slower storage, which does the same job. This memory operates at TLC speed levels, around 578MB/s in tests, but it too can be exhausted.

Eventually, if both initial and secondary buffers are full, the data is directly written to the storage. However, this is at the mentioned glacial speed.

The testing also found that the amount of free space on the drive can make things worse, too. With the drive at 60% capacity, the lower amount of free space cuts into the fast cache capacity severely, from about 250GB of SLC cache to nearer 58GB.

This also impacts the TLC buffer, cutting the speed down to 396MB/s. With the buffers exhausted again, the QLC writing speed averages to 45MB/s, and occasionally hit a glacial 1.1MB/s.