📅 04 Nov 2026 • HPCwire / Argonne ALCF
Argonne scales the ALCF Inference Service
Argonne National Laboratory is scaling up the ALCF Inference Service, its shared national inference infrastructure for open science, per the Argonne Leadership Computing Facility (ALCF), announcement dated 4 November 2026 and circulated through the HPCwire off-the-wire channel. The centerpiece is Tara, a new HPE Cray EX4000 system with 672 nodes and 2,688 NVIDIA GH200 chips. Per ALCF, the expansion gives the service roughly 100 times more GPU memory than it had before, and that changes what research teams can do: larger models, longer contexts and heavier agentic loops run without leaving the facility stack.
Mixed silicon in one pool
Per ALCF, Minerva brings 64 NVIDIA B200 accelerators in 8 new nodes, so a second accelerator family joins the pool. The Metis cluster grows from 2 to 6 SambaRacks, for a total of 96 SambaNova SN40L processors. GH200, B200 and SN40L sit side by side inside one inference service, so a project matches a model to the architecture that fits its latency and memory profile instead of forcing every workload through the same hardware.
Agents as the science interface
The service already hosts more than 50 models for open science, per ALCF. The notable shift is the tooling layer: Codex, Claude Code, OpenCode and Pi already run on the service, and scientists use them to launch agentic workflows for real research tasks. The agent plans the steps of an experiment, submits jobs, reads results and iterates, while inference serves the model calls underneath — the terminal agent becomes the primary interface to scientific compute.
Genesis Mission as demand
Per ALCF, the expansion adds capacity for Genesis Mission, the national initiative on AI for science, alongside scientific research across the broader ALCF user community. Capacity is built ahead of a large program and operated as shared infrastructure. The transferable pattern for platform teams: one inference layer serves the whole organization, heterogeneous accelerators sit behind it, agentic CLIs run on top as first-class tools, and a concrete demand stream justifies the investment.
Sources:
- https://www.hpcwire.com/off-the-wire/argonne-adds-ai-inference-capacity-for-genesis-mission-and-scientific-research/
- https://www.alcf.anl.gov/
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
Top comments (0)