National laboratory leverages SambaStack for energy-efficient inference to drive scientific discovery

TL;DR

• Argonne National Laboratory runs SambaStack SN40 for low-latency agentic AI inference, serving roughly 1,000 scientists who use coding agents daily.
• 200 billion tokens per month are consumed by Argonne researchers on agentic workflows, a figure the lab expects to double.
• SambaStack draws an average of 20 kW or less, a more energy-efficient inference option than GPU-based alternatives for workloads that don't require Argonne's 40 MW Aurora supercomputer.
• Argonne has partnered with SambaNova for seven years and deploys SambaStack on-premises within the Argonne Leadership Computing Facility as part of its Metis system.
• SambaNova RDU three-tiered memory (DDR, HBM, SRAM) enables Argonne to run multiple models with fast switching times for Genesis Mission workloads.

Background

Celebrating 80 years of scientific discovery, Argonne National Laboratory is a U.S. Department of Energy (DOE) research center where scientists and engineers work together to solve some of the greatest challenges in the world. Their projects include cutting-edge biomedical research to building safer nuclear reactors, discovering the electronic properties of materials, and designing magnetic fields that can contain plasma in fusion reactions, and more.

The laboratory works with academic institutions, private industry, and other agencies to tackle the biggest challenges, combining AI models with modeling and simulation into complex workflows to dramatically shorten the time to discovery.

The Genesis Mission

Launched in November 2025 under Executive Order 14363, the Genesis Mission is an ambitious national initiative designed to leverage the power of AI to solve some of the most challenging issues facing America today. As part of the Genesis Mission, Argonne National Laboratory has been awarded funding to lead a broad range of AI-driven research projects on some of the nation's most complex challenges and to redefine how scientists harness AI.

The Genesis Mission is the overarching national initiative designed to dramatically accelerate scientific and engineering productivity through advanced AI methods, workflows, models, and tools. It spans the scientific, national security, and energy sectors, and will expand over time to include other agencies, all the national laboratories, private industry, and economic partners.

Argonne Leadership Computing Facility (ALCF)

The ALCF is a first-of-its-kind AI inference service designed to empower scientists across the country to accelerate research and discovery. It provides secure, private cloud access to a range of LLMs and scientific foundation models on the advanced systems at Argonne National Laboratory.

Challenge: Delivering High Performance Inference, Cost Effectively and at Scale

Argonne National Laboratory supports thousands of scientists and engineers across an incredibly diverse range of workloads, consuming billions of tokens every day. They need an efficient, cost-effective method of delivering these tokens to their users.

Delivering on the needs of the Genesis Mission, as well as other initiatives, requires:

• Fast, low-latency inference for coding models. About 1,000 scientists use coding agents every day, not just for coding, but for all kinds of problem solving. Almost all of them use some form of agentic coding, and that number is expected to grow.
• Cost-efficient, persistent agents. Users consume about 200 billion tokens per month on agentic workflows, a figure expected to double in the near future.
• Energy efficiency as a core priority. Argonne runs the Aurora supercomputer, which consumes about 40 MW for its GPUs. They need a more power-efficient solution on a per-inference basis for workloads that don't require a supercomputer.
• Extreme scale for complex workflows. Argonne is moving toward models that can remember days or weeks of work with the same level of accuracy as standard context, driving toward unprecedented context lengths.
• Disaggregated inference. Argonne is interested in separating the prefill and decode phases of inference to accelerate performance and support greater context lengths.

Solution: SambaStack for Energy-Efficient, Low-Latency Inference

Argonne National Laboratory uses SambaStack™ SN40 to power a range of high-throughput, low-latency AI inference workloads that bring AI and HPC together.

Agents and Coding

Approximately 1,000 scientists at Argonne Labs use persistent agents for far more than coding. They rely on agents as personal assistants for everything from writing and reviewing proposals to planning experiments and reading and writing papers. The number of agents in use is growing exponentially, currently around 200 billion tokens a month, a figure expected to double.

SambaStack delivers high-throughput, low-latency inference to researchers at Argonne, and has proven to be an incredibly cost-effective way to produce leading-edge tokens.

High-Efficiency, Low-Energy Consumption

Not every workflow requires a 40 MW supercomputer. Consuming an average of 20 kW or less, SambaStack is a much more efficient inference solution than GPU-based alternatives.

Three-Tiered Memory

SambaStack is powered by the SambaNova Reconfigurable Dataflow Unit (RDU), which features a three-tiered memory design incorporating DDR, HBM, and SRAM. This enables Argonne to run a suite of models with extremely fast switching times.

Small Footprint for On-Premises Deployment

As a national lab, security is paramount for Argonne. Built with Dataflow Architecture, RDUs are far more efficient than alternative systems, giving SambaStack a small footprint for fast, easy on-premises deployment. Argonne runs the SambaStack SN40 within the ALCF as part of its Metis system, which provides secure, private cloud services to researchers.

Why Argonne National Laboratory Continues to Choose SambaStack

Argonne National Laboratory has been a customer of SambaNova for seven years and continues to be a valued partner. The reasons for this include:

1. Energy efficiency. Air-cooled and with lower power consumption than alternatives, SambaStack keeps workloads cost-effective, which is critical for a data center already managing up to 40 MW of supercomputing load.
2. Secure, on-premises deployment. With a small footprint, SambaStack is easily deployable in existing data centers, including air-gapped environments.
3. Fast model switching. The RDU's three-tiered memory design enables Argonne to run multiple models with extremely fast switching times.
4. Low-latency inference. Argonne uses the SambaStack SN40 as part of the Metis system within the ALCF to provide fast, secure inference services to researchers.
5. Broad open-source model support. SambaStack runs both the open-source models and the scientific frontier models that Argonne and the other DOE labs depend on as part of the Genesis Mission.
6. Long-term partnership. As a member of the Genesis Mission Consortium, SambaNova is building on the long-term relationship that exists with Argonne.

The Results

Argonne National Laboratory continues to take advantage of the energy efficient, low latency inference services that SambaStack delivers and is looking forward to continuing to do so in the future as they drive the next generation of scientific discovery.