Supercomputers/Systems/Aurora
USA · operational · mixed
Aurora
Deep dive: Aurora: eight years from contract to first list
Also known as A21
Operated by Argonne National Laboratory at Argonne Leadership Computing Facility .
Researched with AI assistance, checked by scripts. These rows were researched by AI agents reading public operator, laboratory and vendor sources, appended by scripts and checked by automated validation; no subject-matter expert has reviewed them (how this site is made). A mechanical second-reader pass has since fetched every cited source: 1350 of 1581 systems have a readable citation that names them, and 64 are genuinely weakly sourced. Every claim carries its source and a confidence tier. Treat anything below verified as a lead, not a citation.
Measured performance
- Rmax
- 1.01 EFlop/s
- Rpeak
- 1.98 EFlop/s
- Rmax ÷ Rpeak
- 51.1%
- Cores
- 9,264,128
- Power
- 38.70 MW
- Per watt
- 26.2 GF/W
Figures are as last publicly reported for the configuration described below, not a live measurement. Where a system was upgraded in place, the post-upgrade configuration is shown and the earlier one appears in the timeline.
How it compares
Ranked against the other machines in this dataset that report the same figure: those that first came into service within two years of Aurora, and those in USA. Computed from our tables; a metric is ranked only when at least eight peers report it.
| Metric | Value | Against its era | Against its country |
|---|---|---|---|
| Rmax | 1.01 EFlop/s | #3 of 257 · top 1% | #3 of 332 · top 1% |
| Energy efficiency | 26.2 GF/W | #80 of 135 · bottom 42% | #25 of 162 · top 15% |
| Rmax ÷ Rpeak | 51.1 % | #242 of 256 · bottom 6% | #309 of 331 · bottom 8% |
| Power draw | 38,698 kW | #2 of 139 · top 1% | #1 of 171 · highest |
Every metric is ranked largest-first, so a high power-draw rank means a big power draw, not a good one. Rmax ÷ Rpeak uses HPL-measured results only. Rankings move as machines are added; they describe this dataset, not the world.
What this machine is made of
One row per supplier relationship. “Supplier at build” is the company that shipped the part at the time; where that company has since been acquired, the parent it rolls up to today is shown beside it. That distinction is what makes ticker-level aggregation possible across a thirty-year dataset.
| Role | Part | Supplier at build | Quantity | Confidence | Source |
|---|---|---|---|---|---|
| Integrator | - HPE Cray EX | Hewlett Packard Enterprise | - | Reported | alcf.anl.govintel.com |
| Accelerator | Intel Data Center GPU Max 1550 Accelerators oneAPI (Xe-HPC) · Intel 7 / TSMC N5 · 600 W Intel Data Center GPU Max, 6 per node | Intel | 63,744 reported | Reported | alcf.anl.govintel.com |
| CPU | Intel Xeon CPU Max 9470 CPUs x86-64 · 52C · 2 GHz · Intel 7 · 350 W Xeon CPU Max 9470 52C, 2 per node | Intel | 21,248 reported | Reported | alcf.anl.govintel.com |
| Interconnect | HPE Slingshot-11 Interconnect Dragonfly · 200 Gb/s per port · Around 2 microseconds Slingshot-11 | Hewlett Packard Enterprise | - | Reported | alcf.anl.govintel.com |
| Storage | Intel DAOS Storage Object store, no POSIX layer underneath · 230 PB on Aurora · About 31 TB/s DAOS, 230 PB | Intel | - | Reported | alcf.anl.govintel.com |
| Cooling | Direct liquid cooling, warm water Cooling Direct-to-chip cold plates, facility water loop · Removes 90+ percent of node heat to liquid direct liquid cooled cabinet | Hewlett Packard Enterprise | - | Reported | alcf.anl.govintel.com |
| Operating system | HPE Cray Operating System Operating systems SUSE-derived Linux, container-based management · Proprietary layer over open-source Linux HPE Cray Operating System | Hewlett Packard Enterprise | - | Reported | alcf.anl.govintel.com |
Engineering, funding and access
What Aurora runs on, who paid for it and how people get time on it. Each line carries who said it and when.
Hardware
| Item | Detail | As of | Basis | Source |
|---|---|---|---|---|
| Compute nodes | 10,624 nodes21,248 Xeon Max CPUs and 63,744 Data Center GPU Max GPUs, 2 CPUs and 6 GPUs per node, 166 compute groups | 2025 | Operator stated | arxiv.org |
| Total memory | 10,620 TBDDR5 only (10.62 PB); the paper also lists 9.52 PB of HBM2e across the CPUs and GPUs | 2025 | Operator stated | arxiv.org |
| Storage capacity | 220 PBDAOS tier of 1,024 servers: about 260 PB raw, about 220 PB with 16+2 erasure coding; separate roughly 100 PB Lustre system | 2025 | Operator stated | arxiv.org |
| Interconnect and topology | HPE Slingshot-11 single-dimension dragonfly, 175 groups (166 compute, 8 storage, 1 service), 2.12 PB/s injection bandwidth | 2025 | Operator stated | arxiv.orgtop500.org |
Software
| Item | Detail | As of | Basis | Source |
|---|---|---|---|---|
| Parallel filesystem | DAOS (about 220 PB usable, 31 TB/s peak) plus a roughly 100 PB Lustre file system named Flare with 100 gateway nodes to DAOSDAOS ranked first on the IO500 production list at SC23 per the paper | 2025 | Operator stated | arxiv.org |
| Job scheduler | PBSThe paper mentions PBS job arrays for ensembles; system scheduler is not otherwise described there | 2025 | Operator stated | arxiv.org |
| Programming and software stack | Intel oneAPI compilers (icx, icpx, ifx) with SYCL, OpenMP 5 offload and OpenCL 3.0 over the Level Zero runtime | 2025 | Operator stated | arxiv.org |
| Programming and software stack | MPICH with PMIx and HPE Cray PALS launch; oneMKL, oneCCL and oneDAL librariesTOP500 lists aurora-mpich 51.2 and MKL 2023 update 2 | 2025 | Operator stated | arxiv.org |
| Programming and software stack | Kokkos, RAJA, chipStar and OCCA portability layers; PyTorch, TensorFlow, Horovod, DeepSpeed and JAX for AI | 2025 | Operator stated | arxiv.org |
Funding
| Item | Detail | As of | Basis | Source |
|---|---|---|---|---|
| Funded by | U.S. Department of EnergyExascale Computing Project application development ended in 2023 | 2025 | Operator stated | arxiv.org |
Access
| Item | Detail | As of | Basis | Source |
|---|---|---|---|---|
| How researchers get time | INCITE and ALCC allocations, with ESP early-science teams in the first campaignFirst production allocation period had 34 INCITE and 13 ALCC projects | 2025 | Operator stated | arxiv.org |
Results
| Item | Detail | As of | Basis | Source |
|---|---|---|---|---|
| Notable result | HPL-MxP mixed-precision benchmark of 11.64 exaflops on 9,500 nodes, first place at SC24 | 2024-11 | Operator stated | arxiv.org |
| Notable result | HPL of 1.012 exaflops on 9,234 nodes, third place on the TOP500 at SC24 | 2024-11 | Operator stated | arxiv.orgtop500.org |
AI datacenter metrics
Figures beyond an accelerator count, each with who stated it. What can and cannot be compared across AI datacenters.
| Metric | Value | As of | Basis | Note | Source |
|---|---|---|---|---|---|
| Accelerators | 63,744 accelerators | 2023-07 | Operator stated | Argonne release: 10,624 blades, 6 Intel Max GPUs each (product computed); the second page states 63,744 directly. | news.uchicago.eduglennklockwood.com |
| Stated investment | 0.5 USD bn | 2019-03 | Operator stated | DOE 2019 announcement: contract valued at over $500M (Intel prime, Cray subcontractor). Original contract value, not final cost; excludes facility. | energy.gov |
Published benchmark results
- HPL-MxP 2025-11: HPL-MxP mixed-precision (rank 2), 11.643 EFlop/s (independent, 2025-11) hpl-mxp.orgtop500.org
- HPL-MxP 2026-06: HPL-MxP mixed-precision (rank 2), 11.6 EFlop/s (independent, 2026-06) top500.orgtop500.org
The read
The longest gap between announcement and first light of any system in this dataset: contracted in 2015 around a Xeon Phi design that was cancelled, redesigned around Ponte Vecchio, and delivered eight years later. That lag is the single data point most responsible for the shape of our announcement-to-first-light chart, and it is a vendor-execution story rather than a facility one. Two Xeon Max CPUs and six Data Center GPU Max accelerators per node, on HPE Slingshot-11: a rare case of the accelerator and integrator roles sitting with different companies at this scale.
Timeline
- 2015 Announced source
- 2023-11 List debut at #2 (November 2023) source
- 2023-12 First operational source
- 2024-11 Ranked first on the HPL-MxP mixed-precision benchmark at 11.64 exaflops and third on the TOP500 at 1.012 exaflops at SC24 source
- 2025-01-27 Entered production and general availability, with 16 early-science projects scaled to at least 2,000 nodes beforehand source
Public list appearances
Pointers only: rank and edition, linking to the canonical entry. We do not reproduce list tables. See sourcing policy.
Site & facility
Argonne Leadership Computing Facility
- Location
- Lemont, IL, USA
- Cooling
- Direct liquid cooling (warm water)
- Commissioned
- 2006
Change history
- 2026-10-04 AI benchmark record: hpl-mxp 2025-11
- 2026-10-04 AI benchmark record: hpl-mxp 2026-06
- 2026-09-22 Gained a Green500 rank pointer (#37 on the 2023-11 list)
Nearby in scale
The machines closest to Aurora by Rmax, smaller to larger.
Related systems
Same site
- Sunspot · USA, operational, 2022
- Equinox · USA, planned
- Polaris · USA, operational, 2021
- Solstice · USA, planned
- Theta · USA, decommissioned, 2017
- Cooley · USA, operational, 2015
Same operator
- Improv · USA, operational, 2023
- BeBop · USA, operational, 2017
- Blues · USA, operational, 2012
- Cetus · USA, operational, 2012
- Mira · USA, decommissioned, 2012
- Vesta (Argonne) · USA, operational, 2012
Same accelerator
- Lichtenberg II Phase 2 · Germany, operational, 2023
- Dawn · United Kingdom, operational, 2024
- Stampede3 · USA, operational, 2024
- Clementina XXI · Argentina, operational
- SuperMUC-NG Phase 2 · Germany, operational
Source check
We fetched this system's own citations and recorded whether each page actually mentions it. This is a corroboration signal, not a fact check, and it is published so you can see how well the sourcing holds up rather than take it on trust.
One cited source was readable and does not mention this system; another could not be read because the host returned a bot challenge. The unreadable one may well be fine, so this is weaker evidence than a corroborated row and much stronger than an unsupported one. See how this check works.
| Cited source | Result | Found on the page |
|---|---|---|
| alcf.anl.gov | could not be fetched | - |
| intel.com | does not name this system | Intel Data Center GPU Max 1550, Direct liquid cooling, warm water, Intel Xeon CPU Max 9470 |
Checked 2026-10-18 by pnpm verify. Re-run it and the table changes with the web.
Related news
- Argonne's Aurora Supercomputer Powers 2 of the Largest Cosmological Simulations to Date · HPCwireTwo cosmology runs used most of Aurora's nodes for several days, a rare public data point on how much of an exascale machine a single production job can occupy. It speaks to Aurora's usable scale in practice rather than its peak rating.
- TOP500's 66th list confirms JUPITER as the first exascale system outside the US · TOP500El Capitan holds number one at a remeasured 1.809 exaflop/s and Frontier and Aurora hold two and three, so all three US exascale systems in this dataset keep their relative order. JUPITER's 1.000 exaflop/s Booster partition becomes the fourth exascale system worldwide and the first in Europe, confirming the milestone this dataset's earlier rows on Bull's JUPITER delivery had been building toward.
- Aurora enters service at Argonne · Argonne Leadership Computing FacilityEight years after contract award, following the cancellation of its original Xeon Phi design and a redesign around Ponte Vecchio. The longest announcement-to-first-light gap in this dataset.
Further reading
Inside a node
Why accelerators exist, what the memory hierarchy costs, and how the CPU and accelerator merged onto one package.
Power and cooling
Why air ran out at 20 kW a rack, how warm-water cooling removes the chillers, and what PUE conceals.
Operating a supercomputer
Acceptance testing, why "installed" and "in production" are six months apart, and the machine lifecycle.