[numpad0]

Maybe latency. IIRC flash is a lot laggier than DRAMs and SRAMs.

[DoctorOetker]

The random access memory models is not really representative of ML workloads (both training and inference), where multiplying large tensors result in predictable memory access patterns.