Written and maintained by CASRAI Editorial Board
Published
On this page
The frontier supercomputer, located at the Oak Ridge Leadership Computing Facility (OLCF) at Oak Ridge National Laboratory (ORNL), represents a major breakthrough in high-performance computational science. Commissioned by the U.S. Department of Energy (DOE) Office of Science and built by Hewlett Packard Enterprise (HPE), Frontier achieved international recognition in May 2022 as the world’s first verified exascale system. Operating above 1.1 exaflops (1.1 quintillion calculations per second), Frontier delivers massive computational capability to advance open science across astrophysics, climate modeling, nuclear physics, and artificial intelligence.
Executing scientific workloads at exascale requires researchers to navigate structured competitive allocation programs, optimize software for AMD GPU hardware, and orchestrate data across high-performance parallel file systems. This guide details Frontier’s hardware architecture, scientific scope, allocation pathways, Slurm batch workflows, and leadership-class alternatives.
What it is
Frontier is an exascale supercomputer built on the HPE Cray EX liquid-cooled architecture. The system executes complex multi-physics simulations and data-intensive pipelines that exceed the capabilities of previous petascale supercomputers. In May 2022, Frontier reached a verified benchmark of 1.102 exaflops on the High-Performance Linpack (HPL) test, formally breaking the exascale computing threshold.
The physical system comprises 74 HPE Cray EX cabinets installed at the OLCF data center in Oak Ridge, Tennessee. Each cabinet weighs roughly 8,000 pounds and utilizes direct liquid cooling to dissipate thermal output directly from processors and accelerators. Under peak computational workloads, the facility draws approximately 24.6 megawatts (MW) of electrical power.
Who runs it and what it covers
Frontier is owned by the U.S. Department of Energy and operated by UT-Battelle, LLC, for Oak Ridge National Laboratory under the DOE Office of Science Advanced Scientific Computing Research (ASCR) program. Compute allocations serve researchers across universities, national laboratories, and private industry.
Frontier’s computational scope spans five primary scientific domains:
- Astrophysics and Cosmology: Modeling stellar nucleosynthesis, core-collapse supernovae, and dark matter distribution across cosmic scales.
- Climate and Earth Science: Running global cloud-resolving climate models, such as the DOE Energy Exascale Earth System Model (E3SM), to project long-term climate dynamics.
- Energy and Nuclear Engineering: Simulating whole-core reactor fission physics and plasma turbulence in magnetic confinement fusion experiments.
- Materials Science and Chemistry: Executing ab initio quantum molecular dynamics to discover novel catalysts and battery electrolytes.
- Genomics and Scientific AI: Training foundation models and executing high-throughput macromolecular structure prediction across shared research data collections.
This workload is powered by 9,856 compute blades configured as follows:
- Host Processor: One 64-core AMD EPYC 7A53 (“Trento”) CPU per node running at a 2.0 GHz base clock (up to 3.7 GHz boost), supporting 128 hardware threads.
- Host Memory: 512 GB of DDR4-3200 memory per node, delivering ~205 GB/s peak bandwidth.
- GPU Accelerators: Four AMD Instinct MI250X accelerators per node. Each MI250X package contains two Graphics Compute Dies (GCDs), providing eight logical GPUs per compute node.
- GPU Memory: 64 GB of High-Bandwidth Memory (HBM2E) per GCD (512 GB total HBM2E per node), providing 3.2 TB/s peak bandwidth per accelerator.
- Node Interconnect: Four HPE Slingshot-11 NICs per node providing 800 Gbps (100 GB/s) aggregate injection bandwidth with adaptive packet routing.
- Local Storage: Two 2 TB NVMe SSDs per node for local scratch caching.
- Center-Wide Storage (Orion): Connected to Orion, a center-wide Lustre parallel filesystem providing ~679 PB usable capacity across NVMe flash and hard drive tiers.
How to search or use it
Researchers access Frontier through competitive allocations, account provisioning, and Slurm job management across six core steps:
- Secure an Allocation: Submit a proposal through the open-science INCITE program (~60% of system time), the mission-driven ALCC program (10% to 30%), or a rolling Director’s Discretion (DD) award for code benchmarking.
- Register User Identity: Establish an account in the myOLCF portal, execute standard DOE user agreements, and complete mandatory cybersecurity training.
- Configure Multi-Factor Authentication: Register an approved RSA SecurID hardware or software authenticator token for cryptographically secured terminal access.
- Connect to Login Nodes: Connect via SSH using your OLCF credentials (
ssh [email protected]) to access compilation, debugging, and job submission utilities. - Stage Project Files to Orion: Navigate to project scratch storage (
$PROJWORK/<project_id>). User home directories ($HOME) are mounted read-only on compute nodes during execution. - Submit a Slurm Batch Script: Load required Cray and AMD ROCm programming environment modules and submit your job script using
sbatch:#!/bin/bash #SBATCH -A <project_id> #SBATCH -J exascale_job #SBATCH -o %x-%j.out #SBATCH -t 01:00:00 #SBATCH -p batch #SBATCH -N 2 #SBATCH --ntasks-per-node=8 #SBATCH --cpus-per-task=7 #SBATCH --gpus-per-task=1 #SBATCH --gpu-bind=closest module reset module load craype-accel-amd-gfx90a module load rocm module load cray-mpich export OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK srun -n 16 ./simulation_binaryMonitor queued jobs with
squeue -u $USERand archive output to long-term storage to adhere to Orion’s 90-day purge policy and your research data lifecycle plan.
Strengths and limits
Deploying large-scale computational campaigns on Frontier offers unmatched throughput while requiring specific architectural considerations:
Key Strengths
- Exascale Performance: Sustained 1.102 exaflops performance enables high-resolution physical modeling without empirical parameter shortcuts.
- High GPU Memory Capacity: 512 GB HBM2E memory per node allows massive simulation grids and deep learning parameter tensors to remain resident on device.
- Fast Interconnect Fabric: The 800 Gbps Slingshot-11 network minimizes latency spikes and buffer congestion across multi-thousand-node parallel communications.
Architectural Limitations
- ROCm Porting Overhead: Codebases written exclusively for NVIDIA CUDA must be ported to AMD ROCm using HIP or OpenMP offload directives.
- Complex NUMA Topologies: Eight GCDs per node require explicit MPI rank and core affinity bindings (
--gpu-bind=closest) to prevent cross-bus latency penalties. - Ephemeral Scratch Storage: Files on Orion inactive for more than 90 days are automatically purged, requiring automated post-run data transfers.
Alternatives
Researchers evaluating leadership computing resources compare Frontier with several peer high-performance installations:
| Supercomputer | Host Facility | Architecture & Accelerators | Linpack Benchmark | Access Model |
|---|---|---|---|---|
| Frontier | Oak Ridge Leadership Computing Facility (ORNL) | AMD EPYC Trento + AMD Instinct MI250X | 1.102 Exaflops | Open science via INCITE, ALCC, and Director’s Discretion |
| Aurora | Argonne Leadership Computing Facility (ALCF) | Intel Xeon CPU Max + Intel Data Center GPU Max | 1.012 Exaflops | Open science and AI research via INCITE and ALCC |
| El Capitan | Lawrence Livermore National Laboratory (LLNL) | AMD EPYC CPUs + AMD Instinct MI300A APUs | >1.7 Exaflops | NNSA national security and stockpile stewardship modeling |
| Summit | Oak Ridge Leadership Computing Facility (ORNL) | IBM POWER9 + NVIDIA Volta V100 | 148.6 Petaflops | Open science, CUDA codebases, and mid-scale simulations |
| Fugaku | RIKEN Center for Computational Science (Japan) | Fujitsu A64FX 48-core ARM Processors | 442 Petaflops | Japanese national research and international collaborations |
Frequently Asked Questions (FAQ)
What makes the Frontier supercomputer an exascale system?
Frontier is an exascale system because it demonstrated a verified computing speed exceeding 1.0 exaflops on the High-Performance Linpack benchmark. In May 2022, it recorded 1.102 exaflops, executing over 1.1 quintillion calculations per second.
How do researchers obtain compute time on Frontier?
Computing time is awarded through three peer-reviewed pathways: the competitive INCITE program (allocating ~60% of facility hours), the mission-aligned ALCC program (allocating 10% to 30%), and Director’s Discretion allocations for code porting and scaling.
Which hardware components power each Frontier compute node?
Each node includes one 64-core AMD EPYC 7A53 Trento CPU with 512 GB DDR4 RAM, four AMD Instinct MI250X GPUs providing eight GCDs with 512 GB total HBM2E memory, two 2 TB NVMe SSDs, and four 200 Gbps HPE Slingshot NICs.
Why does Slurm treat each Frontier node as having eight GPUs?
Each of the four physical AMD Instinct MI250X GPUs contains two independent Graphics Compute Dies (GCDs). The ROCm runtime and Slurm scheduler manage each GCD as an autonomous device, exposing eight logical GPUs per node.
What file system does Frontier use for scratch storage?
Frontier uses Orion, a center-wide Lustre parallel file system providing ~679 PB of usable storage across flash NVMe and hard disk tiers. Orion scratch spaces enforce a 90-day automatic purge policy for inactive files.
Can international researchers apply for time on Frontier?
Yes. The INCITE program is open to research teams from any recognized institution worldwide, provided the project addresses high-impact open-science questions and commits to publishing results in peer-reviewed journals.
Sources
Technical specifications and allocation policies in this guide are verified against official operating documentation:
- Oak Ridge Leadership Computing Facility (OLCF): “Frontier User Guide: System Specifications, Compiling, and Running Jobs” (docs.olcf.ornl.gov/systems/frontier_user_guide.html). Verified: 2026-11-10.
- Oak Ridge National Laboratory (ORNL): “Frontier: The World’s First Exascale Supercomputer Overview & Specifications” (olcf.ornl.gov/frontier/). Verified: 2026-11-10.
- DOE Leadership Computing: “INCITE and ALCC Allocation Programs and Proposal Policies” (doeleadershipcomputing.org/) and OLCF Allocation Guidelines (olcf.ornl.gov/for-users/documents-forms/policies/allocation-policy/). Verified: 2026-11-10.
- OLCF Storage Documentation: “Orion Lustre Parallel File System Architecture and Policies” (docs.olcf.ornl.gov/data/filesystems.html). Verified: 2026-11-10.







