AMD MEXT Memory Management Solution Uses AI to "Lighten the Load" on Memory: DRAM Usage Halved, 99% of Performance Retained

According to reports, AMD released MEXT predictive memory software at its In Depth technical forum. It uses AI to manage data placement between DRAM and SSD, temporarily staging less-used data on flash storage, and then uses AI to predict subsequent demand and move it back to DRAM in advance, reducing the time applications wait for data. Tests of specific agent workloads demonstrated by the company showed that DRAM usage could be halved while still retaining 99% of the performance of an all-DRAM configuration.

AI Predicts Data Demand, SSD and DRAM Divide Work Dynamically

AMD's technical explanation notes that flash access can be about 500 times slower than DRAM, so using SSDs directly as general system storage would inevitably hurt performance. MEXT's approach is to use AI to analyze the access patterns of memory pages, predict which data will soon be used again, and load it back from SSD to DRAM in advance, while data that has not been accessed for a long time is moved to flash to free up DRAM space.

This architecture is a form of memory tiering and differs from simply using the SSD as a swap file. MEXT aims to proactively predict data demand rather than passively swapping data out only when DRAM runs short. The overall process can be simplified as follows: continuously monitor the access frequency and usage patterns of data pages, let AI determine which data may soon be used again, move data pages that are temporarily not needed out of DRAM, and put hot data back into DRAM as much as possible before the program needs it again.

MEXT can be paired with dedicated NVMe SSDs and used in on-premises, cloud, virtual machine, and container environments. Existing EPYC servers can also adopt it.

DRAM Usage Halved, Some Workloads Still Maintain 99% Performance

The real value of MEXT still comes down to workload testing. According to AMD's published comprehensive tests of five types of agent workloads, when using half the DRAM with the rest supplemented by MEXT, performance reaches 99% of an all-DRAM configuration; agent workloads with higher storage demands retain about 90% of performance, indicating that the savings still depend on application characteristics.

In Agentic AI tests, AMD also noted that a configuration of 50% DRAM and 50% MEXT can deliver a 70% performance improvement for Agent usage scenarios with high memory demands. If DRAM is reduced to 25%, the performance/cost ratio can increase to 3.5 times in certain memory-intensive Agent workloads.

In chip design applications, AMD tested with Synopsys PrimeTime and retained about 95% of performance after halving DRAM usage. Additionally, when running EDA work on Azure HX series, expanding effective capacity through MEXT allowed more work to be scheduled simultaneously, with throughput per hour in the test increasing by about 40%.

It should be noted that the relevant data are specific test results published by AMD, and not all server workloads can achieve the same performance. MEXT uses AI prediction and data tiering to let NVMe SSDs handle low-frequency access data, indeed offering another approach to expanding storage capacity. But this does not mean SSDs are about to replace DRAM. The two still differ significantly in latency and bandwidth, and whether MEXT can maintain performance depends on whether AI can accurately predict data demand and whether workloads can tolerate some data being placed on a slower storage tier.

AMD Deploys 145 Servers Itself, Plans to Expand to 2,000

AMD first applied the technology to its own IT environment. Based on an initial deployment of 145 servers, it is estimated that compared with an equivalent storage expansion solution involving additional DRAM purchases, about $4.9 million in DRAM procurement spending can be avoided, reducing the cost of the expansion solution by about 83%, and the company plans to expand to 2,000 servers by 2027.