{"id":"supercomputing-atomistic-history","title":"Atomic-Scale Simulation on Supercomputers","subtitle":"Machine history and capex per PFLOP, and what four decades of hardware bought atomistic simulation.","category":"references","tags":["literature-review","savings-stack","supercomputing","history","hpc"],"source":"articles/lit-review/supercomputing-atomistic-history.md","lang":"en","words":1595,"readMinutes":7,"toc":[{"depth":2,"text":"1. Machine table","id":"1-machine-table"},{"depth":2,"text":"2. Workload evolution (with citations)","id":"2-workload-evolution-with-citations"},{"depth":2,"text":"3. Timeline (compressed)","id":"3-timeline-compressed"},{"depth":2,"text":"4. Cost & access notes","id":"4-cost-access-notes"},{"depth":2,"text":"Items flagged as uncertain or unverifiable","id":"items-flagged-as-uncertain-or-unverifiable"}],"html":"
\n

Provenance: explore agent agent-11 (director-commissioned deep research, backdrop research, 2026-07-21) — materialized verbatim. Quantitative claims are as reported by the research agent from sources it accessed; see citations inline. Citation-verification pass pending before any external publication.

\n
\n

Atomic-Scale Simulation on Supercomputers: Expert Digest

All figures below are sourced; derived numbers are labeled [derived], and anything not cleanly verifiable is flagged [uncertain]. Machine performance numbers are TOP500 HPL (Rmax) unless noted.

\n
\n

1. Machine table

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
MachineSite / debutPeak / HPL (FP64)ArchitectureCostPowerNotes for atomistic sim
Tianhe-2NSCC Guangzhou; #1 June 2013 – Nov 201554.9 PF peak / 33.86 PF HPL16,000 nodes × (2× Intel Xeon E5-2692 + 3× Xeon Phi 31S1P), 3.12M cores¥2.4B ≈ $390M (Wikipedia/Dongarra report)17.6 MW; 24 MW with cooling2015 US export ban blocked its upgrade → rebuilt as Tianhe-2A with domestic Matrix-2000 (61.4 PF HPL)
Sunway TaihuLightNSCC Wuxi; #1 June 2016 – June 2018125.4 PF peak / 93.01 PF HPL40,960× SW26010 (260 cores @1.45 GHz), 10.65M cores, no accelerators (Dongarra report, Klockwood notes)¥1.8B ≈ $270–273M incl. R&D (DCD, SCIO)15.37 MW (HPL)First all-domestic Chinese #1; later a proving ground for ML-potential MD (see §2)
SummitORNL OLCF; #1 June 2018 – June 2020200.8 PF peak / 148.6 PF HPL4,608 nodes × (2× IBM POWER9 22c + 6× NVIDIA V100)~$200M (Oak Ridge Today)~13 MW full system (Next Platform); HPL power ~10 MW [uncertain]Machine that launched MLIP-at-scale: 2020 Gordon Bell DPMD run; 3.3 mixed-precision "exaops" advertised
SierraLLNL (NNSA); online Oct 2018, #2 Nov 2018125.7 PF peak / 94.6 PF HPL4,320 nodes × (2× POWER9 + 4× V100), 17,280 GPUs (LLNL, ACM ICS'25)Not separately disclosed [uncertain] (CORAL program total ~$525M, FedScoop)~11–13 MW [uncertain]Classified stockpile stewardship; decommissioned (LLNL page)
FugakuRIKEN R-CCS Kobe; #1 June 2020 – June 2022537 PF peak / 442 PF HPL158,976× Fujitsu A64FX (48-core Arm SVE + HBM2), 7.63M cores, no GPUs (DCD)Reported ~$1B total program (CACM) [reported, not official]~30 MW (Fujitsu spec: 30–40 MW, Fujitsu)First machine to break 2 EF on HPL-AI (mixed precision); shared use from 9 Mar 2021
LUMICSC Kajaani, Finland; full system Nov 2022~530–552 PF peak [uncertain, sources differ] / 379.7 PF HPLLUMI-G: 2,560 nodes × (1× AMD "Trento" 64c + 4× MI250X); HPE Cray EX (CSC/ScienceBusiness, RankRed)€200M lifecycle (EuroHPC; total ecosystem budget >€600M per CSC)7.1 MW, hydro-powered with waste-heat reuse (DCD)Europe's flagship GPU machine; major MLIP/materials user community
FrontierORNL OLCF; first exascale, #1 June 2022 – Nov 20241.685 EF peak / 1.102 EF HPL (June 2022); 1.353 EF (Nov 2024)9,408 nodes × (1× AMD EPYC "Trento" 64c + 4× MI250X), HPE Cray EX235a (arXiv survey)$600M (widely reported)21.1 MW during HPL run (OSTI/IEEE paper); 52.2 GF/WHost of 2023 and 2024 Gordon Bell atomistic prizes (§2)
AuroraANL ALCF; exascale verified May 20241.98 EF peak / 1.012 EF HPL (on 87% of system)HPE Cray EX; Intel Xeon Max 9470 + Intel Data Center GPU Max ("Ponte Vecchio"), 6 GPUs/node (HPE)$500M+ contract (DOE, 2019)38.7 MW (Wikipedia)10.6 EF on HPL-MxP (mixed precision) — largest AI-capable system at delivery
El CapitanLLNL (NNSA); #1 since Nov 2024; dedicated 9 Jan 20252.746 EF peak / 1.742 EF HPL11,136 nodes × 4× AMD MI300A APU (44,544 APUs; 24 Zen4 cores + 228 CDNA3 CUs each), 5.4 PB HBM3 (LLNL, chinaaet/Top500 report)$600M (Caliber.Az/AP)~30 MW; 58.89 GF/W Green500 (IT Pro)Classified NNSA stockpile work; unclassified sibling Tuolumne exists
\n

2. Workload evolution (with citations)

Phase A — plane-wave/pseudopotential DFT era (–2015). Community codes VASP and Quantum ESPRESSO (Giannozzi et al., JPCM 21, 395502, 2009; cited in QE exascale papers) dominated leadership allocations; routine DFT topped out at ~10³ atoms. Linear-scaling efforts (PEXSI, ~10–20k atoms; Lin Lin slides) were the scale frontier. This era maps to Tianhe-2's Xeon+Phi architecture.

\n

Phase B — GPU adoption + classical MD at scale (2016–2020). QE's exascale rewrite is documented in "Quantum ESPRESSO toward the exascale" (Giannozzi et al., JCP 152, 154105, 2020) and the follow-up JCTC 19, 6992 (2023). LAMMPS + Kokkos became the portability vehicle; DOE's Exascale Computing Project funded the EXAALT (Exascale Atomistics for Accuracy, Length, and Time) project, whose SNAP kernel re-engineering gave a 5× speedup on Summit (DOE FY2021 budget justification).

\n

Phase C — MLIP era (2018–2022).

\n\n

Phase D — exascale + foundation models (2022–now).

\n\n

AI-for-science programs. DOE's "AI for Science, Energy and Security" report (May 2023, ANL/ASCR overview); the FASST (Frontiers in AI for Science, Security and Technology) initiative (ANL/HPC User Forum deck). Atomistic-specific allocations: 2025 INCITE project "exascale ab initio reactive MD (AIRMD)" awarded 380,000 Frontier node-hours (2025 INCITE fact sheets); 2026 INCITE includes an "Ab Initio Foundation Model" MLIP project (2026 fact sheets).

\n
\n

3. Timeline (compressed)

\n
\n

4. Cost & access notes

Capex per peak PFLOP [derived — capex ÷ Rpeak]:

\n\n

Effective /FLOP−hour[derivedestimate,electricityonly]:∗∗at/FLOP-hour [derived estimate, electricity only]:** at/FLOP−hour[derivedestimate,electricityonly]:∗∗at0.05–0.10/kWh, Frontier's 21.1 MW costs ~1,050–2,100permachine−hour;dividingbyits1.353EFHPLrategivesroughly∗∗1,050–2,100 per machine-hour; dividing by its 1.353 EF HPL rate gives roughly **1,050–2,100permachine−hour;dividingbyits1.353EFHPLrategivesroughly∗∗800–1,600 per exaflop-hour of delivered Linpack work. Real applications sustain far below HPL, so effective cost per useful FLOP is typically 3–10× higher [rough estimate — electricity price and duty cycle assumptions, not an official figure]. Amortizing Frontier's 600Mcapexovera5−yearlifeadds 600M capex over a 5-year life adds ~600Mcapexovera5−yearlifeadds 14k/machine-hour (~$10k per HPL-exaflop-hour) [derived].

\n

Access models:

\n\n
\n

Items flagged as uncertain or unverifiable

\n"}