this is discussion of amd advertising, so grain of salt and all of that, blah blah about them stopping starving processors. the interesting part is on page 12 about mext

The premise starts with how badly DRAM is used. AMD cited published work from Meta and Microsoft showing that 50-80% of memory pages tend to go cold, 25% of total DRAM is stranded, and half of all VMs never access 50% of the DRAM assigned to them. Meanwhile DRAM has become the most expensive component in a server.

MEXT’s answer is prediction. A Linux kernel driver transparently moves cold pages to flash, and an AI engine predicts which pages will be needed soon and pulls them back into DRAM before the access happens. AMD describes the engine as an ensemble of models including transformer-based LLMs, trained the same way a language model predicts the next word, except the tokens are page accesses.

EDA—Synopsys PrimeTime held 95% of performance with half the DRAM removed (1.9x perf per DRAM dollar), 90% at 62.5% removed (2.4x), and 80% at 75% removed (3.2x). Spark—a TPC-DS-derived SQL workload on a three-node Azure cluster held 99% of performance at 50% DRAM plus 50% MEXT. Throughput—a four-job PrimeTime run on Azure HX-series finished in 6.6 hours against a 9.3-hour baseline, using 2.6 TB of effective capacity against 1.4 TB of real DRAM.

maybe mr. ram will become cheaper when corpos buy amd (or nand becomes more expensive, because those operations probably chew through resources, like those macbooks)