AMD Acquires ‘Predictive Memory’ Startup
AMD acquired MEXT, a developer of AI-driven memory optimization technology
By Mark LaPedus
AMD this week acquired MEXT, a startup that is developing AI-driven memory optimization technology.
The deal will bring new capabilities to AMD, a processor supplier based in Santa Clara, Calif. Founded in 2023, MEXT recently released its first product, dubbed Predictive Memory. MEXT’s technology is a software-only solution that makes flash appear as DRAM to the operating system.
It addresses a major problem in today’s systems--DRAM. DRAM, which is used as main memory in systems, is fast but it’s also a volatile memory type. That means the data is lost in the device when the system is shut down. Then, when a system is running, a DRAM is refreshed every few milliseconds to retain data. This increases power consumption and latency in systems.
“For modern workloads, performance is increasingly constrained not by CPU speeds, but by DRAM capacity,” according to MEXT, a startup based in Santa Clara, Calif. “On top of that, DRAM is one of the most expensive computing resources. For such a costly resource, the expectation would be that it would be well-utilized. In many environments, however, that is not the case.”
Studies from cloud providers have shown that utilization often drops to 50% or lower—over half of the memory can be considered cold, according to the company. This means that organizations are deploying much larger DRAM footprints than they truly need—resulting in potentially millions of dollars in wasted spend, according to MEXT.
MEXT’s Predictive Memory product line attacks the DRAM problem by transparently bringing flash into the memory domain. “Much like how large language models predict the next logical word in a natural language pattern, MEXT’s AI engine predicts the next logical memory page in a workload’s behavioral pattern,” according to the company.
According to MEXT, the process works in three steps: MEXT identifies memory pages that are not actively in use (cold) and offloads them to flash, which costs 50x less than DRAM. MEXT’s patent-pending AI engine predicts which offloaded memory pages will be needed soon. The engine proactively pushes those pages back to DRAM before they are required, so the application experiences little to no performance impact.
“Modern data center infrastructure is evolving rapidly, and customers are increasingly facing a common challenge: access to memory,” said Dan McNamara, senior vice president and general manager of Compute and Enterprise AI at AMD, in a blog. “As AI models, data analytics, virtualization and high-performance computing workloads grow in size and complexity, memory has become a critical constraint across cloud and enterprise environments. For customers, addressing these bottlenecks is essential to improving performance per dollar, increasing efficiency and accelerating deployments at scale.”
MEXT’s approach, according to AMD, has the potential to reduce infrastructure costs, improve resource utilization, and help customers more effectively scale general-purpose and AI workloads.
The acquisition expands the AMD’s portfolio and helps customers with memory optimization technology designed to improve performance, reduce total cost of ownership and accelerate time to deployment.