AMD Says Its MEXT Software Can Cut Server DRAM Needs by Up to 75 Percent While Keeping 95-99 Percent of Performance

AMD has disclosed internal deployment results for MEXT, a predictive memory-tiering technology it acquired in June 2026, presenting them at the company's In Depth technical forum. The pitch is blunt: most server workloads are paying for DRAM they barely use, and a small AI model can decide what actually needs to sit in memory. According to AMD's figures, the software has already cut an estimated $4.9 million in DRAM purchases across the first 145 internal servers, with memory expansion costs 83 percent lower than an equivalent all-DRAM build.

The mechanism is closer to a smarter swap file than a hardware change. MEXT moves infrequently accessed "cold" pages from DRAM to NVMe flash storage, then uses AI to predict which pages an application will touch next and pulls them back to memory before they are needed. The operating system still sees the full memory capacity it was promised, so applications run unmodified. The stack ships as a Linux kernel module that works with any distribution, plus an AI engine and a monitoring tool called RAM View.

The AI itself is deliberately modest: a transformer-based language model that trains in under a second on first deployment, then keeps learning each workload's access patterns. AMD says it runs on a single EPYC core with no GPU required. Because page ordering does not affect results, the model only has to guarantee that a page is back in DRAM by the time it is requested — a looser constraint than exact prediction, which is what makes the approach work at all.

The company's reported benchmarks are the strongest claim in the announcement. On a three-node Azure cluster of fifth-generation EPYC processors running Spark TPC-DS, a configuration with half the DRAM retained 99 percent of full-DRAM performance. In Synopsys PrimeTime, a memory-hungry chip-design tool, half the memory kept 95 percent. In a geometric-mean test with five AI agents running concurrently, half memory kept 99 percent and a quarter kept 85 percent. AMD says DRAM demand can fall 20 to 75 percent depending on the workload, and that for multi-agent and memory-intensive applications, a 50/50 DRAM-flash split can raise performance per dollar of DRAM by 70 to 90 percent. All of these are company-run tests, not independent measurements.

The economics explain why AMD cares. Gary Smerdon, MEXT's former CEO and now a corporate vice president at AMD, has said memory can account for 60 to 90 percent of the cost of a large server. With DRAM prices climbing on AI demand and supply tight, replacing part of that DRAM with cheap NVMe flash changes the math on every rack. The catch is physics: flash can be hundreds of times slower than DRAM, so the whole scheme rests on the predictor rarely being wrong.

Deployment is moving in phases. The first 145 internal servers already run MEXT, with plans to scale to 2,000 servers by 2027. AMD is partnering with HPE to bring the technology to enterprise customers, and early hardware is already visible: some EPYC Venice servers now carry "MEXT Ready" branding with dedicated NVMe bays. Licensing comes in the second phase, and AMD is starting with its own EDA workloads — chip design tools being exactly the kind of memory-hungry, repetitive-access application the approach favors.

If the numbers hold up under independent testing, the implications reach past AMD. Agentic AI workloads multiply memory appetite across orchestration, sandboxing and data staging, and every hyperscaler is currently paying spot prices for the DRAM to feed them. A software layer that cuts DRAM needs by half at near-zero performance cost would soften demand at the margin — and undercut the pricing power that memory makers have enjoyed through the current shortage. For now, every performance figure in this story carries the same caveat: they are AMD's own.