New Memory System Cuts Bottlenecks in Data Centers

Representational image of data centers

This post is also available in: עברית (Hebrew)

Data centers today face a growing challenge: modern servers must juggle multiple programs from countless users simultaneously, all while keeping data secure. This creates bottlenecks, slows performance, and leaves much of the hardware underused. Now, an international team at EPFL has introduced a new approach to virtual memory called Midgard, aimed at overcoming these limitations.

Virtual memory allows a program to operate as if it has more memory than physically available, managing multiple applications securely. Traditional systems, however, were designed decades ago and are not optimized for today’s high-demand cloud services. Each request undergoes strict identity and access checks, which ensure security but add delays and increase energy use.

Midgard addresses this by reorganizing virtual memory into compartments, enabling programs to locate their workspace quickly while hardware performs access checks more efficiently. This redesign preserves the necessary security protocols but significantly reduces memory bottlenecks. Importantly, it works with existing software standards, so developers don’t need to adapt their applications.

According to TechXplore, the benefits extend beyond performance. Data centers often operate with unused memory—reports indicate that more than half of rented memory is idle, and Microsoft has noted up to 20% of memory can remain unallocated in certain cloud setups. By improving memory management, Midgard not only accelerates servers but also increases overall hardware utilization, which can reduce energy consumption and help lower the carbon footprint of data centers.

Over the past five years, the Midgard project has produced numerous publications, a demonstrator showing substantial efficiency gains, and presentations to industry leaders like Intel. Looking forward, the team plans to expand its concepts to rack-scale computing, where an entire rack of servers operates as a single system. Coupled with emerging high-speed network fabrics, this approach could allow cloud services to use all available hardware more effectively—not just memory, but processors and other resources as well.

Midgard represents a step toward a more efficient, sustainable, and high-performance future for data centers, rethinking core principles of computing to meet the demands of modern cloud infrastructure.