{"id":"savings-dft-systems","title":"Systems-Level DFT Acceleration","subtitle":"LS-DFT, GPU ports, mixed precision and reduced-rank methods, SCF accelerators, k-point economization, and wavefunction reuse.","category":"references","tags":["literature-review","savings-stack","dft","systems","hpc"],"source":"articles/lit-review/savings-dft-systems.md","lang":"en","words":2852,"readMinutes":13,"toc":[{"depth":2,"text":"1. Executive result","id":"1-executive-result"},{"depth":2,"text":"2. Savings table","id":"2-savings-table"},{"depth":2,"text":"3. Proven vs claimed","id":"3-proven-vs-claimed"},{"depth":2,"text":"4. Openings for Lupine","id":"4-openings-for-lupine"}],"html":"
\n

Provenance: explore agent agent-17 (director-commissioned deep research, swarm of 7, 2026-07-21) — materialized verbatim, then editorially corrected only to replace the retracted 624/132 union-anchor figures with the reproducible primary-record values (558/154) and update arithmetic derived from them. Quantitative literature 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

Chapter digest: Systems-level DFT acceleration (LS-DFT, GPU ports, mixed precision / reduced-rank, SCF accelerators, k-point economization, wfn/density reuse, ML guesses)

Evidence cut: 2026-07-21. Every number below was read this session; each citation is tagged [full text] or [abstract only]. Items I could not verify are marked [UNVERIFIED].

\n

1. Executive result

\n

2. Savings 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\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\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
TechniqueWhat cost it removesMeasured savings (system / hardware context)Accuracy cost / failure modeCitationEvidence
Linear-scaling DFT (CONQUEST)O(N³) diagonalization; enables 10⁶-atom DFTMillions of atoms LS; multi-site support functions: 4 vs 17 functions/atom (4× fewer) with exponentially converging energy error; Sakurai–Sugiura eigensolve on 194,573-atom Ge/Si: 2,399 s on 6,400 K-computer nodesTruncated density-matrix range; on 8-atom cell LS solver is 2× slower than TZTP diagonalization (crossover); support-function/basis restrictions for sparse inverse overlapNakata et al., JCP 152, 164112 (2020), arXiv:2002.07704v3 [full text]strong
Linear-scaling DFT (ONETEP, BigDFT, CP2K)Same"Thousands of atoms" LS at PW accuracy (ONETEP); "many thousands of atoms" LS (BigDFT); CP2K aimed at "massively-parallel and linear-scaling"Localization truncation; metals need finite-temperature tricksPrentice 2020, DOI 10.1063/5.0004445 [abstract only]; Ratcliff 2020, DOI 10.1063/5.0004792 [abstract only]; Kühne 2020, arXiv:2003.03868 [full text, partial]moderate
CP2K ADMM + HFX screeningExact-exchange cost in hybrid DFTScreening O(N⁴)→O(N²); ADMM brings HFX "to within a few times the cost of conventional GGA"; molopt basis trim = 10× faster density mapping for liquid water; LRIGPW: 64-water box mapped with 192 atom terms instead of ~200,000 pair terms (99% distant pairs)ADMM is an approximation (auxiliary basis); operator truncation radius system-dependentKühne et al., JCP 152, 194103 (2020), DOI 10.1063/5.0007045 [full text]strong
GPU port, Quantum ESPRESSOCPU→GPU per-node speed1.4× at 98 atoms → >3× for larger systems (1×V100 vs 18-core Xeon E5-2697); earlier CUDA-Fortran port 2–3×; numerics identical (ΔE ≤ 2·10⁻⁸ Ry)Small systems don't saturate GPU; memory per card is the bottleneckGiannozzi et al., JCP 152, 154105 (2020), arXiv:2104.10502v2 [full text]strong
GPU port, VASP (OpenACC)Same; also fixes CUDA-C maintainabilityOpenACC port matches or beats the CUDA-C port; clear advantage at 8 GPUs from better scaling (VASP 5.4.4, V100/P100 era); community benchmarks 2.9–8.8× (2×V100 vs 8 CPU nodes)RMM-DIIS vs Davidson coverage; tiny jobs unaccelerated; NSC slide numbers not peer-reviewedMaintz & Wetzstein, CUG 2018 proc. [full text]; NSC workshop slidesmoderate
GPU port, FHI-aims (all-electron NAO)Hamiltonian integration, density update, forces on GPU2.4–6.6× for key steps; 3–4× full calculations over a 103-material test set; near-ideal scaling on 375-atom Bi₂Se₃Only real-space ops offloaded; eigensolver (ELPA) separateHuhn et al., Comput. Phys. Commun. 254, 107314 (2020), DOI 10.1016/j.cpc.2020.107314 [abstract + description in arXiv:2409.09399]moderate
GPU port, ABACUS (LCAO)Grid integrals + diagonalization on GPUDemonstrated up to 10,444 atoms (twisted bilayer graphene); grid integration time ~linear: 2.41 s @ 76 atoms → 93.95 s @ 2,524 atoms; bottleneck migrates to eigensolver (76% of SCF time at 1,876 atoms)End-to-end speedup factors are in the results section I could not extract from HTML [UNVERIFIED: exact × number]; accuracy unchanged (same math)Zhang et al., arXiv:2409.09399v3 [full text, methods]moderate
GPU port, SPARC real-space (semilocal & hybrid)CPU node-hours & core-hoursHybrid: up to 8× node-hours / 80× core-hours, ~300 s for >6,000-electron metallic system (V100); semilocal predecessor: up to 6× node-hours / 60× core-hours, <30 s for >14,000 electronsKronecker-solver batching; ACE needed for exchangeJing, Sharma, Pask, Suryanarayana, arXiv:2501.16572 [abstract]; semilocal 2023 numbers as quoted in arXiv:2501.16572 intro and arXiv:2409.09399 [secondary]strong (hybrid) / moderate (semilocal, secondary)
GPU port, DFT-FE (finite element)All key kernels on GPU~20× CPU→GPU node speedup; 80–140 s ground states for 6k–15k-electron benchmarks; up to ~100,000 electronsHigher-order FE meshes; GPU memory ceilingDas et al., arXiv:2203.07820v2 (DFT-FE 1.0; CPC 2022) [abstract]strong
GPAW GPU (CuPy)Real-space kernels on GPUGPU support "achieved with minor modifications of the GPAW code thanks to CuPy" — no speedup numbers in abstract [UNVERIFIED: quantitative gain]—Mortensen et al., GPAW review, arXiv:2310.14776v2, DOI 10.1063/5.0182685 [abstract]claim
Reduced-rank exchange: ACEExchange-operator application & update countPer-SCF-iteration cost "only marginally larger than GGA"; "orders of magnitude speedup for Hartree-Fock-like calculations"; advantageous already at tens of atomsOne-time ACE construction overhead; gap-independentLin, JCTC 12, 2242 (2016), arXiv:1601.07159v2 [abstract]strong (widely replicated in use)
Reduced-rank exchange: ISDFNumber of Poisson solves per exchange applicationO(N_e²)→O(N_e) Poisson solves; ~2 orders-of-magnitude exchange-cost reduction; 1,000-atom bulk-Si HSE converged in 10 min on 2,000 cores; scales to 8,192 cores at 4,096 atomsISDF interpolation error (controlled by rank parameter); metallic cases need careHu, Lin, Yang, JCTC 13, 5420 (2017), arXiv:1707.09141v1 [abstract]strong
Mixed/single precision + dual grid in hybrid PW DFTFP64 FFT and ISDF/ACE cost"Several times" further acceleration with "acceptable trade-off in accuracy"; single-precision FFTs alone ≈2× on FFT-bound parts; 8,000 Si-atom hybrid on 16,000 CPU coresAccuracy trade-off is explicit but bounded; DP retained for non-HFX partsHou et al., JCTC 21, 787–802 (2025), DOI 10.1021/acs.jctc.4c01541 [abstract]; 2× FFT detail via arXiv:2601.08077 backgroundmoderate
Mixed precision in QC kernels (TeraChem lineage; DFT-FE)FP64 tensor/GEMM costTeraChem mixed SP/DP: ~2× on Kepler-era GPUs at near-DP accuracy; DFT-FE used mixed precision to reach 33% of benchmark performance on SummitLegacy-hardware dependent (SP:DP throughput ratio); final iteration in DP recommendedMixed-precision survey, arXiv:2412.19322 [full-text section read]moderate
Periodic Pulay mixingSCF iterations3× fewer iterations than DIIS (Pd-bulk); Fe-bulk @1000 K: 52 → 17 iterations (k=2); converges at 25 K where Pulay fails (>250 iter); 2–3× over Pulay in an independent spectral code (ClusterES)Extra parameter k; gains largest for metals/low TBanerjee, Suryanarayana, Pask, CPL 647, 31–35 (2016), arXiv:1512.01604v2, DOI 10.1016/j.cplett.2016.01.033 [full text]strong
Predictive (trust-region) mixingSCF iterations; robustnessPolyak-step estimate from history; tested on 8 structures × 4 mixing schemes + 36 fixed cases; "works well independent of candidate step"Few absolute iteration numbers in abstract [moderate detail]Woods et al., arXiv:2104.04384 [abstract]claim
Wfn/density reuse in AIMD: XL-BOMD / CP2G ASPCIterative ground-state optimization per MD stepXL-BOMD: "without the requirement of an iterative, non-linear electronic ground state optimization prior to the force evaluations and without a systematic drift in the total energy"; CP2G: "substantially reduce the number of SCF iterations required per MD step" (needs consistent tuning of propagation+corrector+thermostat)Energy drift if mistuned; ASPC not time-reversible (high-order error)Niklasson, JCP 147, 054103 (2017), arXiv:1705.10845, DOI 10.1063/1.4985893 [abstract]; CP2G best-practices, arXiv:2601.12191v2 [abstract]strong (MD) / moderate (CP2G tuning guide)
Wfn/density reuse along NEB bandsFirst-SCF cost per image per NEB iteration[UNVERIFIED] — standard practice in VASP/QE (restart WFC per image), but I found no 2015–2026 paper quantifying the savings along a band——[UNVERIFIED]
k-point GR gridsNumber of irreducible k-points at fixed accuracyMetals at 1 meV/atom: 60% faster than Monkhorst-Pack, 20% faster than simultaneously-commensurate grids (>7,000 structures tested); finer density incrementsErratic convergence for metals remains; benefit mainly metalsMorgan et al., Comput. Mater. Sci. 153, 424 (2018), arXiv:1804.04741v2 [abstract]moderate (single group, large test set)
Smearing economizationk-points needed for metal convergence[UNVERIFIED] — no dedicated quantified 2015–2026 study found; treated implicitly inside GR-grid and mixing papers——[UNVERIFIED]
ML density-matrix/density initial guess (molecules)Early SCF iterations−33.3% SCF iterations, trained on ≤20-atom molecules, applied to 60-atom and 900-atom systems; Hamiltonian-prediction baselines +80% iterations or divergenceMolecular (Gaussian-basis) scope; periodic extension openLiu et al. (ByteDance Seed), arXiv:2509.25724v3 [abstract]moderate
ML periodic density guess (ELECTRAFI)Early SCF iterations + total DFT costUp to ~20% total DFT compute reduction for crystalline materials; 633× faster inference than strongest prior density model (sub-second); slower density models negate savingsSavings capped by SCF-iteration share; plane-wave codes only so farÆrtebjerg et al., arXiv:2601.19966v2 [full text]moderate (very new, single group)
ML spectral guess for BigDFT (LimitX)Early SCF iterations at scaleChebyshev-spectrum prediction "bypasses early SCF iterations in BigDFT"; 2 TB protein-dimer datasetNo wall-clock numbers yet; project-stagearXiv:2606.00401 [abstract via search]claim
DGDFT + PEXSI (adaptive local basis)Degrees of freedom & diagonalization~15 ALBs/atom at planewave accuracy (vs hundreds of PW coefficients); PEXSI avoids diagonalization; phosphorene ribbons to ~10,000 atomsO(N²) PEXSI cost in 3D; can't exploit good initial guesses between geometry stepsHu, Lin, Yang, JCP 143, 124110 (2015), arXiv:1506.08147 [abstract]strong
PWDFT-SW (Sunway)CPU heterogeneity exploitation64.8× speedup at 4,096 Si atoms; up to 16,384 C atoms (as described in the ABACUS GPU paper's related-work section; original report not directly accessed)Sunway-specificvia arXiv:2409.09399 [secondary]claim
PWmat (GPU-native PW, China)CPU→GPUVendor materials claim ">30× vs CPU version" via mixed precision + GPU-native design; peer-reviewed basis: Jia et al., CPC 211, 8 (2017) [reference seen, abstract not read]Marketing numbers; no system-size/error accounting in blog sourcesvendor/press; CPC ref via [arXiv:2501.16572 bibliography]marketing

3. Proven vs claimed

Replicated and trusted

\n\n

Single-paper / single-group claims

\n\n

Marketing / non-citable

\n\n

Thin sub-areas (honest assessment): ONETEP has no dedicated peer-reviewed GPU-port paper in 2015–2026 that I could find; k-point smearing economization and wavefunction reuse along NEB bands have no quantified literature — both are practitioner knowledge, marked [UNVERIFIED].

\n

4. Openings for Lupine

\n

Sources accessed this session (all links above): 25+ primary papers — full text read for CONQUEST (arXiv:2002.07704v3), QE-exascale (arXiv:2104.10502v2), Periodic Pulay (arXiv:1512.01604v2), CP2K (arXiv:2003.03868, partial), ABACUS GPU (arXiv:2409.09399v3, methods), VASP OpenACC CUG 2018 (partial), ELECTRAFI (arXiv:2601.19966v2, partial); abstracts verified for ACE, ACE-ISDF, Hou, DFT-FE, SPARC-hybrid-GPU, ByteDance guess, GR grids, XL-BOMD, CP2G, ONETEP, BigDFT, GPAW, predictive mixing, DGDFT, LimitX, mixed-precision survey. [UNVERIFIED] items: ABACUS GPU end-to-end × (results section not extractable), NEB wavefunction-reuse savings, smearing economization, PWmat numbers (marketing), PWDFT-SW (secondary).

\n"}