AMD buys Mext to ease AI-driven memory constraints

AMD’s latest acquisition targets a bottleneck of its own making: the DRAM shortage fueled by AI workloads. On June 16, the chipmaker announced its purchase of Mext, a predictive memory startup founded in 2023, for an undisclosed sum. The deal aims to integrate Mext’s machine learning-driven memory tiering into AMD’s enterprise and AI compute offerings, potentially reducing reliance on expensive DRAM without sacrificing performance.

What happened

AMD acquired Mext, a startup developing a proactive memory platform that uses AI to dynamically shift data between DRAM and flash storage. Mext’s software, which runs as a daemon (Mextd), analyzes access patterns to predict which memory pages are "cold" (infrequently used) and offloads them to flash. When data is needed again, the system preemptively restores it to DRAM, minimizing latency penalties. The approach leverages modern flash arrays, which now rival DRAM in aggregate bandwidth, though not in latency. Mext claims its technology can expand a system’s effective memory capacity by 2 to 4 times using flash, which remains significantly cheaper per gigabyte than DRAM.

Memory tiering is not new—Intel’s Optane persistent memory and earlier software-based solutions have explored similar concepts. Mext distinguishes itself by using a combination of heuristics, long short-term memory (LSTM) models, and transformer architectures to optimize data migration, akin to a branch predictor for memory. AMD’s Dan McNamara, SVP of compute and enterprise AI, stated the acquisition could "lower infrastructure expenses, boost resource efficiency, and enable better scaling for both traditional and AI workloads."

"This approach has the potential to reduce infrastructure costs, improve resource utilization, and help customers more effectively scale general-purpose and AI workloads." — Dan McNamara, SVP of AMD’s compute and enterprise AI business (via blog post)

Why it matters

The DRAM shortage, driven largely by demand for AI training and inference, has strained data center budgets and limited the scalability of large language models (LLMs). Mext’s technology could mitigate these constraints by allowing enterprises to deploy larger models with fewer high-bandwidth memory (HBM) resources. For example, mixture-of-experts (MoE) models, which route tokens to specialized sub-models, often underutilize some experts. AMD may use Mext’s predictive algorithms to offload infrequently used experts from HBM to slower system memory, reducing costs without degrading performance.

Beyond AI, the acquisition could benefit general-purpose workloads in cloud and enterprise environments. By expanding effective memory capacity, Mext’s platform may delay hardware upgrades or reduce the need for additional servers, particularly in memory-constrained environments like virtualized infrastructure or high-performance computing (HPC).

What to watch

AMD has not disclosed a timeline for integrating Mext’s technology into its products, but the startup’s software-based approach could accelerate deployment. The success of the integration will depend on minimizing latency overhead and ensuring compatibility with existing AMD hardware, including EPYC CPUs and Instinct accelerators. Competitors like NVIDIA and Intel may also explore similar memory-tiering solutions if Mext’s approach proves effective, potentially leading to broader adoption of AI-driven memory optimization in data centers.