Back to the portfolio

Automatic workload placement across heterogeneous memory tiers

Computing, Software & AIElectronics & SemiconductorsConceptPatent pending

This technology creates a smart memory management system that watches how a processor is actually performing and uses those real-time observations to predict what will happen if it mixes two different types of memory together. Based on those predictions, it automatically sets rules for how the two memory types should be used together. The workload then runs under those rules, balancing performance across the memory mix without requiring manual tuning. In practical terms, this means a server or computing system can dynamically optimize how it uses, for example, fast DRAM alongside slower but larger persistent or tiered memory, extracting better overall performance.

What you could build

A firmware or OS-level memory tiering module that could be embedded in cloud server platforms or data center operating systems to automatically optimize workload placement across heterogeneous memory (e.g., DRAM + CXL-attached or persistent memory); primary buyers would be hyperscalers, cloud infrastructure vendors, and server OEMs.

Who in Virginia should care

Northern Virginia's massive data center corridor — home to AWS, Microsoft Azure, Google, and dozens of colocation operators — makes this directly relevant to any hyperscaler or server hardware vendor with a local presence.

Readiness: Concept

Concept — described but not yet demonstrated. Lab validated — supported by experimental results in the patent. Prototype likely — the text describes a built, working embodiment.

Readiness is inferred from the patent text, not from a lab visit.

The record

Inventors
YOUNGSAM SHIN, JINSHU LIU, HAMID HADIAN, YEONGON CHO, HUAICHENG LI, CHANGWOO MIN
Filed
Patent pending — filed January 14, 2025
Status
Application
Publication number
US20250321792A1

Ready to talk?

Virginia Tech Intellectual Properties handles licensing for this technology.

VTIP contact coming shortly

Prosim summaries are generated from public patent text and are not legal advice.