A unique active memory computer purpose-built for AI science applications

August 21, 2025 by Karyn Hede, Pacific Northwest National Laboratory

Collected at: https://techxplore.com/news/2025-08-unique-memory-purpose-built-ai.html

With the particular needs of scientists and engineers in mind, researchers at the Department of Energy’s Pacific Northwest National Laboratory have co-designed with Micron a new hardware-software architecture purpose-built for science.

The Crete prototype system will offer 15 terabytes of active memory co-located with the system processors. This memory-to-processor configuration is unique within the DOE laboratory complex and even among the high-performance computing community at large, according to system designers.

It came online August 12, 2025, for DOE Office of Science national laboratory users, who can apply through the Advanced Memory to Support Artificial Intelligence for Science (AMAIS) initiative for access to this one-of-a-kind testbed.

“We are exploring novel memory technologies to address applications and problems that are constrained by memory shortcomings that manifest in the form of capacity, system bandwidth or sharing,” said Andrés Márquez, the AMAIS initiative lead and a PNNL computer scientist who specializes in designing new computing technologies.

“We are using Compute Express Link (CXL), an open industry standard, to connect memory with the processor via an I/O switch developed by XConn Technologies.”

This platform expands on existing architectures where memory is, by design, tightly coupled to one or more processing elements. The Crete system provides a clever mix of two types of memory architectures—tightly coupled, which is supported by Micron Registered Dual In-line Memory Module, as well as loosely coupled via custom Micron CXL controller boards—offering a tremendous opportunity for data-driven scientific computing, including AI, according to Márquez.

Crete’s random-access memory equals the memory in 240 high-end laptops running with 64 gigabytes of memory if they were all running simultaneously and communicating in real time.

“The Crete system represents a bold leap forward in redefining how memory and compute can work to unlock scientific discovery,” said Mark Helm, senior fellow of Micron’s Pathfinding and Strategy team in the Technology and Products group.

“At Micron, we believe that the future of AI-driven research depends on architectures that break free from current memory constraints. Our collaboration with PNNL is not just in the development of a powerful system—it’s about enabling a new era of exploration, where memory-rich environments empower scientists to uncover deeper insights to solve critical problems.”

The new Crete testbed is available to be used by researchers from other national laboratories and academia who have applications that require more memory than is available on most of today’s computing architectures, which tend to favor processing speed over large memory access.

“For several decades, the HPC community has focused on distributed memory system architectures, and while many applications can be partitioned to make use of memory that is directly attached to processors, it has led to ‘orphaned’ applications that require large directly addressable memory capacity,” said James “Jim” Ang, chief scientist for computing at PNNL.

These chemistry applications that integrate AI are the first candidates for users on Crete.

“We have computational chemistry applications that are going to use this larger memory space right away,” added Ang. “We’re excited to find out what Crete can do with generative AI analyzing scientific data, as opposed to large language models and text.”

A memory chat room

The new memory architecture is designed to offer a space where large databases, molecular modeling algorithms and AI agents can meet, like a chat room, to exchange information in real time and make predictions, suggest new research paths and analyze experimental results.

“The Crete testbed will allow us to push the leading edge of what CXL technology is capable of and to define what scientific AI looks like for the coming decade,” said Márquez.

GenAI algorithms are advancing daily, and researchers need to be able to integrate AI with their simulation and data analysis workflows. Crete will provide a launchpad for a new generation of algorithms that accelerate discovery in chemistry, advanced materials and molecular biology.

AMAISing AI for science

The AMAIS initiative supports a number of scientific and cybersecurity projects to advance AI for science.

“Building upon the innovative vision of the late computer scientist Jim Gray, who introduced the concept of “The Fourth Paradigm: Data-intensive Scientific Discovery,” Crete aims to showcase the potential of this idea by accelerating large-scale, data-intensive, memory-starved scientific workloads that can benefit from AI,” said Márquez, principal investigator of the AMAIS project.

“The memory mechanism enabled by the XConn memory switch allows memory that is sitting outside in the network to be easily accessed by multiple hosts at the same time, passing information without going to the hard drive for large workloads,” he added. “That’s a big differentiator. Nobody else is doing that.”

Anticipating this hardware-software co-design, PNNL invested in laboratory-directed research funding for Software Defined Architectures, a fully open-source toolchain for highly specialized hardware accelerators focusing on machine learning. That project began several years ago in the Data Model Convergence initiative and now serves as a foundational tool essential in AMAIS.

Researchers also plan to perform ongoing analysis of how the new memory system performs by assessing the major bottleneck of HPC: data access from the memory system. The research team, led by PNNL’s Nathan Tallent, will deploy purpose-built memory analysis tools that deliver detailed insight without requiring excessive additional time, space and resources.

The team has also built in defenses to protect the system from cyberattack or other cybersecurity issues.

“We know that there are a lot of cybersecurity vulnerabilities when it comes to memory,” said Márquez. “That tends to be the Achilles heel of any type of security system in the compute world. We are addressing cybersecurity vulnerabilities in the design phase, right from the get-go, into the system.”

Ready for exascale

The Crete testbed is designed for complex computing workflows with a system architecture that can be reconfigured as needs arise.

For example, a team of PNNL researchers has partnered closely with Micron to leverage the emerging CXL technology for computational chemistry. During the development of Crete, computer scientist Ajay Panyala and computational chemist Karol Kowalski worked with the Micron team to enable the exascale-ready computational chemistry software ExaChem to run simulations using CXL memory on Crete.

“We are building the tools that anticipate the next-generation flagship supercomputers commissioned by DOE user facilities,” added Ang. “This investment bridges the gap between today’s computing infrastructure and those coming online in the next decade.”