AMD Acquires MEXT to Obtain Memory Tiering Technology

AMD announced on June 15 (local time) that it has acquired MEXT, a startup specializing in memory tiering technology. MEXT’s “Predictive Memory Engine” enables NAND flash storage to be recognized as DRAM by operating systems and applications, potentially reducing the need for expensive DRAM in data centers. The terms of the acquisition have not been disclosed.

Addressing Memory Constraints in Data Centers

The proliferation of large-scale AI models has led to a rapid surge in memory demand within data centers. According to a report by Tom’s Hardware, memory resources, rather than CPU or GPU processing power, are increasingly becoming the performance bottleneck for entire systems. At the same time, DRAM is often used inefficiently in many environments, with data that is rarely accessed lingering in DRAM, contributing to cost pressures.

MEXT’s technology addresses this challenge with AI-based memory tiering.

How Predictive Memory Engine Works

MEXT’s Predictive Memory Engine continuously analyzes system memory access patterns. Using AI models, it predicts which memory pages stored on NAND flash will be needed next. These predicted pages are proactively transferred to DRAM before being requested by applications. This allows applications to perceive data as if it resides in main memory, maintaining performance while reducing DRAM usage.

NAND flash is significantly cheaper per unit of capacity compared to DRAM. According to AMD, this technology boosts the effective memory capacity of existing infrastructure while reducing dependency on expensive DRAM, contributing to lower total cost of ownership (TCO) for cloud providers and enterprise customers.

AMD’s Data Center Strategy

With this acquisition, AMD plans to integrate MEXT’s technology into its portfolio of data center products. AMD already offers integrated solutions combining CPUs, GPUs, accelerators, networking technologies, and software, and the Predictive Memory Engine will complement this portfolio.

In AI workloads, which demand substantial memory resources, efficient access to large memory pools is key to performance and scalability. AMD intends to extend MEXT’s technology to cater to these workloads. Additionally, the acquisition brings onboard MEXT’s expertise in memory architecture, infrastructure software, and large-scale computing systems.

Editorial Opinion

In the short term, AMD is likely to deploy reference architectures combining MEXT’s technology with its EPYC processors and Instinct accelerators by the end of the year. For cloud providers, DRAM costs make up a significant portion of operational expenses, and if this technology proves effective, it could become a differentiating factor against competing platforms from Intel and NVIDIA. Announcements of proof-of-concept trials by major hyperscalers are expected within the next three to six months.

From a long-term perspective, memory tiering technologies can be seen as a step toward blurring the lines between DRAM and storage. Over the next one to three years, the integration of Compute Express Link (CXL) and other interconnect technologies could standardize more flexible memory configurations. However, challenges remain, including the accuracy of AI-based predictions and latency overhead, which could become barriers in practical implementation. AMD will need to quickly release benchmarks and use cases to address these concerns.

An open question from the editorial team is the extent to which this technology impacts the write endurance of NAND flash. Heavy reliance on NAND as a DRAM alternative could accelerate wear and reduce its lifespan. Enterprise users will need to carefully evaluate the trade-offs between performance gains and storage durability.

References

• AMD takes over MEXT to ‘address growing memory constraints’ in the data center — memory tiering technology enables flash to appear as DRAM to applications - Tom’s Hardware — Published on June 15, 2026