AMD's MEXT cuts server DRAM requirements in half

A memory trick against the DRAM crunch

As AI investment drives up demand for memory chips, DRAM costs are climbing across both consumer and enterprise hardware. AMD has shared more detail on MEXT, its predictive memory engine built to ease that pressure.

How it works

MEXT uses a transformer model and clever use of flash storage to aim at keeping roughly 95% of a server's performance with half the DRAM capacity. In effect, the system predicts which data should stay in fast memory to make up for the missing capacity.

Already in the field

According to AMD, the system is already running on 145 servers. The company plans to expand that to 2,000 servers in 2027.

Why it matters

The memory shortage directly raises data center buildout costs. Approaches like MEXT promise to run the same workloads with less DRAM, cutting both cost and supply pressure.