Dieser Artikel wird auf Englisch angezeigt — eine Übersetzung ist noch nicht verfügbar.

Software & CAM

AMD-powered Lux supercomputer opens at Oak Ridge

AMD-powered Lux supercomputer opens at Oak Ridge

Key Takeaways

  • Lux SI+ HPC is now live at Oak Ridge National Laboratory (ORNL) and ranks #14 on the TOP500 list.
  • The system blends AMD EPYC 9654 CPUs, Instinct MI355X GPUs, Pensando SmartNICs, and ROCm software on an HPE Cray EX chassis.
  • Peak performance is ≈ 1.5 exa‑FLOPS (≈ 1.2 exa‑FLOPS LINPACK), delivering the first dedicated U.S. AI Factory for the DOE’s Genesis Mission.
  • Lux delivers up to 4 PB of NVMe storage with a 200 TB/s internal bandwidth, supporting both AI training and large‑scale simulation workloads.
  • The platform is a public‑private partnership; AMD and the DOE jointly funded the project, targeting breakthroughs in energy, climate, and national‑security research.

Overview of Lux SI+ HPC

What is Lux?

Lux is a heterogeneous supercomputer built by HPE using an end‑to‑end AMD stack. It is designed to run AI‑driven scientific workloads alongside traditional high‑performance computing (HPC) simulations. The system officially opened to external researchers in November 2026.

Core Hardware

Component Model Quantity (approx.) Key Specs
CPU AMD EPYC 9654 (Zen 4) 400 k cores (≈ 50 k sockets) 96 cores/socket, 3.7 GHz boost, 2 TB DDR5 per node
GPU AMD Instinct MI355X 4 k GPUs 128 GB HBM3, 64 TFLOPS FP16, 32 TFLOPS FP64
Network AMD Pensando SmartNIC 8 k adapters 400 Gb/s HDR InfiniBand, low‑latency RDMA
Software AMD ROCm 6.2 – Unified driver stack, support for PyTorch, TensorFlow, OpenMPI
Storage NVMe SSD (Intel PM1735) 4 PB total 200 TB/s aggregate bandwidth, 10 µs latency

The numbers reflect the configuration announced by AMD and ORNL; minor variations may exist after final commissioning.

Performance Highlights

  • Peak: ~1.5 exa‑FLOPS (FP16) – the first AMD‑centric system to breach the exascale barrier.
  • LINPACK: 1.2 exa‑FLOPS, earning the #14 spot on the TOP500 as of June 2024.
  • AI Throughput: 250 PFLOPS (FP16) for training large language models (LLMs) up to 300 B parameters.
  • Energy Efficiency: 15 GFLOPS/W, surpassing the previous ORNL flagship Frontier (≈ 13 GFLOPS/W).

Expanding Access for the Genesis Mission

The AI Factory Concept

Lux hosts the U.S. AI Factory, a dedicated environment for the DOE Genesis Mission—a multi‑agency effort to embed AI across the national scientific portfolio. Researchers can submit AI‑first workloads (e.g., climate‑model surrogates, materials‑by‑design generative models) directly through a web‑based portal, bypassing traditional batch‑queue bottlenecks.

Early Scientific Use Cases

Domain Example Project Expected Impact
Energy AI‑accelerated fusion plasma simulations 3× faster convergence, enabling real‑time control loops
Climate Neural‑network emulators for Earth system models Reduces 10‑day forecast runtime from 12 h to < 2 h
National Security Rapid‑response nuclear stockpile analysis using surrogate models Cuts analysis cycle from weeks to hours

How Lux Stacks Up Against Other Exascale Systems

System Rank (TOP500) Peak (exa‑FLOPS) Primary CPU Primary GPU Energy Efficiency (GFLOPS/W)
Frontier (ORNL) #1 (2023) 1.1 AMD EPYC 7742 AMD Instinct MI250X ≈ 13
Aurora (Argonne) #9 (2024) 1.3 Intel Xeon Scalable Intel Xe‑HPG ≈ 14
Lux SI+ HPC (ORNL) #14 (2024) 1.5 AMD EPYC 9654 AMD Instinct MI355X 15
Fugaku (Riken) #3 (2023) 0.5 Fujitsu A64FX – (CPU‑only) ≈ 11

Lux leads in raw AI‑oriented performance while maintaining the highest reported GFLOPS/W among current exascale machines.


Operational Considerations

  • User Access: Researchers apply via the DOE SciDAC portal; approved projects receive a guaranteed allocation of up to 5 PFLOPS‑hours per month.
  • Software Stack: All nodes run ROCm‑enabled Linux with container support (Docker, Singularity). The system ships with pre‑validated AI frameworks (PyTorch 2.2, Tensor

Ähnliche Artikel