AMD MEXT Tackles DRAM Supply Crisis With AI-Powered Flash Memory Technology

AMD MEXT Addresses Critical DRAM Supply Crisis

AMD unveiled MEXT, an AI-powered memory solution designed to combat the escalating DRAM supply crisis affecting data centers worldwide2

2

. The predictive memory technology enables flash memory such as NVMe SSDs to function as DRAM tiering, directly addressing RAM supply constraints that have pushed DRAM costs to comprise 75% of total server expenses2

2

. While DRAM prices historically dropped from $50 million per GB in the 1970s to $8 per GB in the 2010s, recent prices have surged beyond $10 per GB and continue rising2

2

.

How AI-Driven Memory Optimization Works

The memory solution employs an AI engine featuring transformer-based LLMs specifically designed for memory page management2

2

. MEXT's AI-driven memory optimization uses predictive algorithms to handle data placement between DRAM and flash storage, pushing "soon-to-be-needed" pages to DRAM while offloading cold data to flash memory as DRAM2

2

. The self-improvement algorithm trains in under one second, runs on a single EPYC core, and requires no GPU, making it remarkably efficient2

2

. This tiering optimization addresses a critical inefficiency: 50-80% of total DRAM in data centers remains underutilized as memory pages "go cold," with Microsoft Azure and META experiencing 50% DRAM overprovisioning2

2

.

Performance Benchmarks Across Workloads

AMD MEXT delivers substantial performance per dollar improvements across diverse applications. In Agentic AI workloads, a 50% DRAM and 50% MEXT configuration achieves a 90% increase in performance per dollar within 5-agent workflows, while a 25% DRAM and 75% MEXT mix nets a 3.4x increase in the same concurrency task2

2

. For EDA/Chip Design using Synopsys PrimeTime on AMD EPYC CPUs, MEXT retains up to 95% of total DRAM performance even with a 50% cut in DDR5, delivering a 1.9x performance per dollar ratio2

2

. On Azure VMs running Spark TPC-DS with 5th Gen EPYC processors, the technology enables up to 99% performance on a 3-node cluster with a 50% DRAM/MEXT split versus all-DRAM baseline configurations2

2

.

Technical Architecture and Components

MEXT comprises three core components designed for seamless integration with AMD EPYC CPUs2

2

. The Linux kernel module is a lightweight, low-overhead driver that works with any Linux OS, providing the foundation for tiering optimization1

1

2

2

. The AI Engine features an ensemble of AI models including transformer-based LLMs that continuously train and infer in real-time2

2

. RAM View serves as an observability platform illustrating real-time memory utilization, helping administrators monitor hot data and cold data distribution2

2

.

Cost Savings and Capacity Expansion

The memory solution offers two primary benefits: reducing DRAM usage to mitigate DRAM costs, and expanding memory capacities at lower expenses2

2

. Physical DRAM paired with MEXT can cut DRAM usage by 20-75% while maintaining similar performance, even when running flash storage as system memory2

2

. When paired with MEXT, physical DRAM can deliver a 3x return in capacity at significantly lower costs than adding traditional DRAM2

2

. AMD projects 83% savings versus an all-DRAM alternative, avoiding $4.9 million in DRAM costs on initial 145 server deployments scaling to 2,000 servers by 20272

2

.

Licensing Model Raises Questions

While AMD typically champions open-source software, MEXT represents a departure from this approach1

1

. The technology is available only through AMD customer and sales channels and is not open-source or publicly available1

1

. AMD announced MEXT is currently in Phase 1, ready for deployment across servers with abundant cores and dedicated NVMe drives or bays2

2

. Phase 2 will make MEXT available to license with RAM View pre-installed2

2

. AMD has hinted at expanding MEXT's application beyond data centers into client segments, suggesting potential consumer-facing implementations2

2

.