PNNL

What is AMAIS?

Advanced Memory to Support Artificial Intelligence for Science (AMAIS) is a project at Pacific Northwest National Laboratory (PNNL) that is broadening the applications of artificial intelligence (AI) in science. Innovations in memory technology are crucial for keeping pace with rapid AI advancements and producing next-generation AI systems that will drive scientific discovery.

The project is sponsored by the Advanced Scientific Computing Research program in the Department of Energy, Office of Science.

Why is memory important for AI?

Stronger memory enhances AI scientific computing performance for several key tasks:

• analyzing large datasets
• scaling up AI systems for application in new scientific contexts
• accelerating AI computation to achieve real-time data processing
• improving the energy efficiency of large-scale advanced computing technologies
• solving memory-bound computational problems.

AI is a rapidly growing research field that is helping to drive discovery and advancements in a wide range of science domains. However, its performance and capabilities hinge on underlying memory systems.

How AMAIS supports AI for science

AMAIS supports AI for science by helping to develop new memory technologies that keep pace with advancements in AI. The new memory systems being developed under AMAIS are needed to bring together first-principles scientific modeling/simulation and AI data-driven science.

New AI technologies require larger memory to achieve the high-volume data processing involved in large-scale AI-driven simulations. Research within AMAIS is helping to meet these needs without sacrificing memory bandwidth, latency, or advanced computing performance.

Current AMAIS projects:

• Crete computing system: Crete is a prototype-switched dance-hall processor/memory architecture that offers 15 terabytes of shared, disaggregated memory with near-memory compute capabilities.
• Fabric-attached memory (FAM): FAM is a disaggregated memory architecture that has widespread applications, and it is a key AMAIS focus area because of its major implications for the future of AI-driven scientific research.

What is fabric-attached memory?

FAM is a type of memory architecture consisting of a mesh-like network fabric that ties together distinct “pools” of physical memory. Unlike other memory architectures, FAM separates memory from computer processors, which improves efficiency by enabling memory sharing among individual computer nodes.

For scientific applications of advanced AI and computing, FAM offers several notable benefits:

• Large datasets: FAM is great for applications involving particularly large datasets that would otherwise be too big to fit into local memory in a single node and that would not easily be partitioned without substantial computational costs.
• Scalable AI: Because memory and processing are separate components in this memory framework, AI systems with FAM architectures can upgrade or scale one component (either memory or processing) without changing the other.
• Cost saving: Advanced AI and high-performance computing systems tend to incur high costs for hardware. FAM can dramatically reduce these costs by enabling targeted upgrades and avoiding the need for customized hardware.

To learn more about the advancements that AMAIS is driving in advanced memory architectures and AI for science, read the latest in AMAIS news. 

Related Projects

TEC⁴

The Transferring Exascale Computational Chemistry to Cloud Computing Environment and Emerging Hardware Technologies (TEC⁴) project’s goal is to accelerate the transition from the basic research associated with developing and implementing electronic structure methods to their widespread use in solving complex challenges in industrial chemical sciences.