AI-Focused Memory Computer Advances Scientific Computing
RICHLAND, Wash. — Researchers at the Department of Energy's Pacific Northwest National Laboratory (PNNL), in collaboration with Micron, have developed a new hardware-software system specifically built for scientific and engineering applications. Named Crete, this prototype features 15 terabytes of active memory placed directly alongside system processors, a configuration unique within DOE laboratories and the wider high-performance computing (HPC) community.
Crete became operational on August 12, 2025, and is accessible to DOE Office of Science national laboratory users via the Advanced Memory to Support Artificial Intelligence for Science (AMAIS) initiative. This testbed aims to support researchers with applications that require more memory than is typical in current computing systems, which often prioritize processing speed over large memory capacity.
Innovative Memory Architecture for AI Science
"We are exploring novel memory technologies to address applications limited by memory size, bandwidth, or sharing," said Andrés Márquez, AMAIS initiative lead and PNNL computer scientist. Crete uses Compute Express Link™ (CXL), an open industry standard that connects memory to processors through an I/O switch developed by XConn Technologies.
The Crete system combines two memory architectures: tightly coupled memory supported by Micron Registered Dual In-line Memory Modules, and loosely coupled memory managed through custom Micron CXL controller boards. This hybrid approach offers significant potential for data-driven scientific computing, including AI applications. To put it in perspective, Crete's random-access memory capacity equals that of 240 high-end laptops with 64 GB memory each, all running simultaneously and communicating in real time.
Mark Helm, senior fellow at Micron, emphasized the system’s potential: "The Crete system redefines how memory and compute can collaborate to accelerate scientific discovery. Our partnership with PNNL focuses on creating memory-rich environments essential for AI-driven research."
Addressing Memory Constraints in Scientific Research
Current HPC systems mainly use distributed memory architectures, which can create challenges for applications requiring large, directly addressable memory spaces. James "Jim" Ang, chief scientist for computing at PNNL, highlighted how chemistry applications integrating AI stand to benefit immediately from Crete's expanded memory. These applications will be among the first to utilize the platform's capabilities.
Crete’s architecture acts like a "memory chat room," enabling large databases, molecular modeling algorithms, and AI agents to exchange data in real time. This setup supports advanced predictions, research suggestions, and experimental data analysis, helping push the boundaries of AI in scientific workflows.
AMAIS Initiative and Scientific AI Development
The AMAIS initiative, which funds Crete, also supports various scientific and cybersecurity projects aimed at advancing AI for science. Building on the concept of "The Fourth Paradigm: Data-intensive Scientific Discovery," the initiative focuses on accelerating data-intensive and memory-demanding scientific workloads that benefit from AI integration.
A key innovation is the XConn memory switch, which allows memory located outside the processor network to be accessed simultaneously by multiple hosts without resorting to slower hard drive storage. This capability sets Crete apart from other systems.
PNNL has also invested in open-source software tools for specialized machine learning hardware accelerators, developed through the Software Defined Architectures project. These tools are critical for AMAIS and enable detailed memory system performance analysis without excessive overhead.
Cybersecurity is another priority. The team designed defenses into Crete from the start to mitigate common vulnerabilities in memory systems, a frequent weak point in computing security.
Preparing for Exascale Computing
Crete’s flexible architecture supports complex computing workflows and can be reconfigured to meet evolving research needs. For example, PNNL researchers collaborated with Micron to leverage CXL memory technology for computational chemistry. They adapted the exascale-ready software ExaChem to run simulations using Crete’s CXL memory.
This effort is part of the Transferring Exascale Computational Chemistry to Cloud Computing Environment and Emerging Hardware Technologies (TEC 4) project within the DOE Accelerate initiative. The project fosters partnerships across national labs, academia, and industry to prepare for the next generation of supercomputing.
Jim Ang noted, "We are building tools that anticipate the flagship supercomputers DOE will commission. This investment bridges today’s computing infrastructure with future systems expected in the next decade."
Access and Opportunities
DOE-funded researchers from other national laboratories and academic institutions can apply for access to Crete through the AMAIS initiative. This unique platform is ideal for scientific applications that require substantial memory resources alongside AI integration.
For those interested in expanding their AI skills to leverage systems like Crete, exploring specialized courses in AI and scientific computing can be beneficial. Resources like Complete AI Training’s courses by job role offer practical learning paths tailored to research professionals.
Your membership also unlocks:
