{"id":"743c70ee-c1d8-40a1-abcd-80789f47293c","arxiv_id":"2502.00152","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Deflated CG with 2048 eigenvectors solves HISQ propagators about 10x faster than standard CG on a 144^3 x 288 lattice at the physical light quark mass, excluding eigenvector setup cost.","lead":"This paper tests deflation, a method that removes the slowest modes of the quark matrix before using the conjugate gradient solver, for HISQ quark propagators in lattice QCD. On a 144^3 x 288 lattice at the physical light quark mass, deflation with 2048 eigenvectors reduces the solve time by about a factor of 10.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 10x speedup is a solve-time result that excludes the eigensolve; without setup-cost or crossover data, the practical claim is unquantified.","rationale":"The reader's weakest assumption correctly identifies the amortized-eigensolve condition as the key boundary on the practical claim. My read of the paper confirms that the reported 10x speedup is internally consistent as a solve-time benchmark: the iteration counts, timings, and node-seconds comparison are coherent, and the authors transparently exclude setup costs (Section 2) and disclose the Frontier regression. The missing piece is quantitative: no eigensolve timing, no I/O timing, and no crossover analysis appear anywhere in the paper, despite the final section stating that setup and solver times will be investigated in future work. Without these, the headline 'factor of 10 speedup' cannot be converted into a wall-clock improvement for a real workflow, and the practical recommendation remains conditional on the number of solves per configuration. The secondary issues noted by the reader—single configuration per ensemble, no error bars, and unreported re-deflation tolerances—are real but less load-bearing; they affect precision and reproducibility, whereas the amortization gap affects whether the central practical claim holds at all. Since the reader already assigned CONDITIONAL for exactly this reason, no verdict change is needed.","tokens_in":8451,"tokens_out":9410,"duration_ms":92830,"concrete_test":"Measure the full eigensolve and I/O cost for 2048 eigenvectors on the 144^3x288 ensemble using QUDA's staggered_eigensolve_test and the PARTFILE read path; combine these with the reported 4.79 s per deflated solve and 47.8 s per undeflated CG solve to compute total wall-clock time per configuration as a function of the number of propagator solves. Report the crossover number of solves at which deflated CG becomes faster, and compare it with typical production counts (e.g., number of quark masses, sources, colors, and spin components). If the crossover exceeds typical counts, the practical recommendation and the 10x claim should be restated as solve-time-only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central practical claim—that deflation improves HISQ propagator solves—rests on the assumption in Section 2 that the eigensolve cost is made negligible by amortization over many solves. All reported timings (Table 2, Fig. 3) are solve times only, explicitly excluding eigensolve, eigenvector I/O, and memory allocation. For the 144^3x288 lattice, the 2048 eigenvectors used in the 10x speedup are generated by thick-restarted Lanczos; the paper provides no timing for this step, no disk-read cost for the roughly 20 TB of single-precision single-parity eigenvectors, and no crossover analysis showing how many propagator solves per configuration are needed before deflated CG beats undeflated CG in wall-clock time. If a production workflow computes only a modest number of propagators per configuration, the 10x solve-time factor may not translate into a comparable speedup, and deflation could even be slower than CG once setup is included. The paper is explicit about this scope, so the issue is not internal inconsistency; it is an unquantified boundary condition on the headline result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This LATTICE2024 proceedings paper reports an empirical comparison of deflated versus undeflated even-odd preconditioned CG for HISQ propagator solves. Using the MILC and QUDA libraries, the authors study four ensembles at fixed physical volume (5.76 fm) with lattice spacings from 0.12 to 0.04 fm and quark masses ranging from the strange mass to the physical light mass. Section 3 introduces a multi-deflation scheme in which single-precision eigenvectors are used and the system is periodically re-deflated, triggered when the residual drops by a factor of 1/delta_tr, in order to avoid the stalling observed with sloppy eigenvectors. The central result, in Table 2 and Fig. 3, is that on the 144^3 x 288 lattice at the physical light quark mass, 2048 deflating eigenvectors reduce the average solve time from 47.8 s to 4.79 s, a factor of 10, and reduce the average CG iteration count from 33,003 to 2,909, with a true-residual stopping criterion of |r|/|b| < 1e-8. The paper also reports node-second cost comparisons and openly discloses a Frontier performance regression that led them to use Perlmutter for the final results. All reported solve times exclude eigensolve, eigenvector I/O, and memory allocation, with the stated assumption that the eigensolve cost is amortized over many propagator solves per configuration.","tokens_in":8600,"tokens_out":8395,"duration_ms":93747,"significance":"If the reported solve-time speedups hold, the paper fills a clear gap in the literature by providing HISQ deflation performance data at the largest lattice sizes currently in use, and it offers a useful baseline for future comparisons with HISQ multigrid. The paper is careful in several respects: stopping criteria are fixed, times are averaged over multiple solves with the first (setup-heavy) solve excluded, the Frontier regression is disclosed, and the node-second analysis addresses part of the fixed-resource criticism. The main limitation is that the end-to-end practical speedup depends on an unmeasured eigensolve and I/O cost; the paper is transparent about this scope, but the phrase 'time to solution' in Section 4 overstates what is actually measured. In my view this is a local wording and framing issue rather than an internal inconsistency, and it does not invalidate the scoped claim that deflation speeds up the iterative solve phase.","major_comments":[],"minor_comments":[{"comment":"The sentence 'We show a factor of 10 speedup of the time to solution on the 144^3 x 288 lattice at the physical light quark mass' should be reworded to 'solve time' or should explicitly repeat the caveat that eigensolve, I/O, and memory allocation are excluded; as written, 'time to solution' conflicts with the definition used in Section 3 and Fig. 3.","section":"Section 4"},{"comment":"The text indicates that Table 2 tests multi-deflation with single-precision eigenvectors, but the table caption does not state the eigenvector precision or the delta_tr values used for the periodic re-deflation; please state in the caption that all deflated runs use single-precision eigenvectors and report the delta_tr values or note that they were tuned separately for each lattice size.","section":"Section 3, Table 2"},{"comment":"In the text describing the right panel, the authors refer to 'the blue residual curve,' but neither the figure nor its caption identifies which curve is blue; please add a legend or explicitly state the correspondence so that the claim about delta_tr = 10^-3 is unambiguous.","section":"Section 3, Fig. 2"},{"comment":"The statement 'Because of the V^2 cost of deflation' would benefit from a one-sentence explanation that the number of eigenvectors scales with volume and each deflation application is O(V), so that readers do not conflate eigensolve scaling with the per-solve deflation cost.","section":"Section 1"},{"comment":"Given that the 2048 single-precision single-parity eigenvectors for the 144^3 x 288 lattice are stated to be about 20 TB, a sentence on how the PARTFILE format is used in the timing runs, and whether disk-read time is included in any reported number, would help readers assess the I/O component of the setup that is excluded from the solve times.","section":"Section 3"},{"comment":"The comparison with the multigrid result of Ref. [14] should state explicitly whether the cited 10x speedup is also a solve-time-only result with the same stopping criterion; otherwise the sentence 'This speedup is comparable to...' invites an end-to-end comparison that the present data do not support.","section":"Section 4"}],"recommendation":"minor_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this one. First, it delivers a clean, internally consistent solve-time benchmark: on the 144^3 x 288 lattice at the physical light quark mass, deflated CG with 2048 eigenvectors cuts the average solve time from 47.8 s to 4.79 s, a factor of 10, with iteration counts dropping from 33,003 to 2,909. Second, the headline 10x is a solve-time figure only; the eigensolve, disk I/O, and memory allocation are excluded, and the paper says so directly. That is the main soft spot, but it is a disclosed scope decision rather than an oversight.\n\nWhat is genuinely new and useful: this is the first deflated HISQ benchmark on a lattice this large and at the physical light quark mass, and the multi-deflation scheme with single-precision eigenvectors is a practical memory hack that works. The paper shows the stalling problem with sloppy eigenvectors and demonstrates that periodic re-deflation fixes it, halving the memory footprint while keeping performance close to double precision. That is a real contribution for practitioners who want to use deflation on memory-bound machines.\n\nThe data are handled honestly. They fixed the stopping criterion, reported iteration counts and wall times, separated solve time from setup time, and disclosed the Frontier regression that forced them to switch to Perlmutter. The node-seconds analysis is a good-faith attempt to address the fixed-resource criticism, and it still shows deflation winning by a factor of four in cost efficiency even when undeflated CG runs on the minimum nodes it needs.\n\nWhere I'd push back on the stress-test note: yes, the crossover point is unquantified, but the paper's stated target is the common workflow of many propagator solves per configuration, and within that scope the result is valid. The bigger real weaknesses are that each ensemble uses a single configuration, so there are no error bars on the speedup, and the empirically tuned re-deflation trigger delta_tr is not fully reported across lattice sizes. Those are worth fixing if this goes beyond proceedings.\n\nOverall, this is a well-scoped performance study, not a deep algorithmic derivation. The math is standard, the measurements are reproducible in principle, and the claims match what was measured. It deserves a serious referee, and I would accept it with minor revisions. For anyone working on HISQ solvers or choosing between deflation and multigrid, it is worth a careful read.","headline":"Solid LATTICE proceedings benchmark showing a real 10x solve-time speedup for deflated HISQ propagators at physical light quark mass, with the caveat that setup costs are explicitly excluded.","tokens_in":9202,"tokens_out":1593,"would_cite":true,"duration_ms":18258,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Deflating the low eigenmodes of the HISQ operator cuts light-quark propagator solve time 10-fold at the physical mass on a 144^3 x 288 lattice.","keywords":["lattice QCD","HISQ fermions","deflation","conjugate gradient","critical slowing down","eigenvalue computation","GPU solver","propagator solve"],"falsifier":"Run a production workflow at the physical light quark mass on the $144^3 \\times 288$ lattice that computes a realistic number of propagators per configuration, and record total wall-clock time including eigenvector generation, disk read, and memory allocation; if the deflated total is not below undeflated CG, the amortization premise fails. A specific crossover test is to find the number of solves per configuration at which the two totals are equal.","tokens_in":8213,"feed_emoji":"⚛️","tokens_out":9264,"duration_ms":81435,"temperature":0.7,"pith_summary":"The paper argues that deflation—using the low eigenmodes of the Dirac operator as an initial guess for the iterative solver—removes most of the critical slowing down that makes light-quark propagator solves expensive. At the physical light quark mass on a $144^3 \\times 288$ lattice, deflated conjugate gradient with 2048 eigenvectors took 4.79 s per solve versus 47.8 s for conventional even-odd preconditioned CG on the same 384 nodes, a 10x speedup. This matters because propagator solves dominate the cost of lattice QCD calculations, and multigrid methods for staggered fermions have been slow to mature. The paper also claims that single-precision eigenvectors with periodic re-deflation match double-precision convergence while halving memory and disk footprint. The authors offer this as a baseline for comparing deflation with multigrid on current large lattices.","feed_headline":"Deflation speeds up light-quark solves by 10x on a 144^3 lattice","feed_subtitle":"Projecting out low quark modes cuts 33,003 CG iterations to 2,909 at the physical light mass.","key_machinery":"The machinery is deflated conjugate gradient on the normal equations. Eigenvectors $v_i$ of the massless squared Dirac operator $\\not{D}^{\\dagger} \\not{D}$ are projected onto the source to form the initial guess $x = \\sum_i v_i \\frac{1}{\\lambda_i} v_i^{\\dagger} \\eta$, so the CG solver works only on the high-mode part of the solution. Multi-deflation repeats this projection during the solve when the residual drops by a factor of $1/\\delta_{tr}$, which keeps single-precision eigenvectors from stalling the residual; the re-deflation threshold was chosen empirically.","core_discovery":"The central claim is a measured factor-of-10 reduction in solve time on the largest lattice in the study: with 2048 eigenvectors of the massless squared Dirac operator, the average time to solve the HISQ propagator equation $M^{\\dagger} M \\psi = M^{\\dagger} \\eta$ at the physical light quark mass on a $144^3 \\times 288$ lattice fell from 47.8 s to 4.79 s, and the iteration count fell from 33,003 to 2,909. A second claim is that these eigenvectors need not be stored in double precision: if the solve is periodically re-deflated with single-precision vectors whenever the residual drops by a chosen factor, convergence matches the double-precision case while using half the memory and disk space. The paper presents this as evidence that deflation is a practical alternative to multigrid for HISQ propagator solves at current lattice sizes, even though deflation shifts, rather than eliminates, critical slowing down.","pith_inferences":["A practical consequence not tested here is that the crossover number of solves per configuration, not the solve-time ratio, determines whether deflation pays off; reporting that crossover would let production teams decide when to use the method.","The empirical choice of the re-deflation threshold suggests a testable extension: a rule relating $\\delta_{tr}$ to the precision of the stored eigenvectors could make the method self-tuning.","The paper notes a performance regression on one of the two large supercomputers used, so the 10x figure may not transfer automatically to every architecture; reproducing the timing on another GPU cluster would test how much of the gain is machine-dependent.","If batched multi-right-hand-side solves and eigenvector compression mature as outlined in the outlook, the amortized setup cost could drop enough to make deflation attractive for moderate numbers of propagator solves as well."],"forward_implications":["On the $144^3 \\times 288$ lattice at the physical light quark mass, deflated CG with 2048 eigenvectors gives a 10x solve-time speedup over conventional even-odd preconditioned CG when both runs use the same resources.","The method is also more cost-efficient per solve: 1839 node-seconds per solve with deflation versus 7914 node-seconds for undeflated CG run on fewer nodes at the light quark mass.","Single-precision eigenvectors with periodic re-deflation converge like double-precision ones, halving the memory and disk footprint, which lets the deflated problem run on half as many nodes.","At the strange quark mass the deflated solve is typically slower than undeflated CG; the speedup appears only near the physical light quark mass.","The measured factor of 10 matches a recent multigrid result for the same lattice size and quark mass, giving a direct performance comparison between the two approaches for HISQ."],"supporting_citations":[{"why":"Earlier numerical experiments on deflation with the HISQ action; this work extends them to larger lattices and physical light quark masses.","marker":"[3]"},{"why":"Develops the technique of computing and deflating eigenvalues while solving multiple right-hand side systems in QCD.","marker":"[2]"},{"why":"Reports a 10x multigrid speedup for the same lattice size and quark mass, providing the comparison point for deflation versus multigrid.","marker":"[14]"},{"why":"Describes the mixed-precision GPU solver machinery, including the reliable-update mechanism used in the CG baseline.","marker":"[15]"},{"why":"Shows a multi-deflation approach for approximate eigenvectors that the paper adapts to single-precision eigenvectors.","marker":"[17]"}],"fun_headline_variants":["Deflation speeds up light-quark solves 10x on 144^3","33k CG iterations drop to 2.9k with deflation","Single-precision deflation halves memory without accuracy loss","Deflation: practical multigrid alternative for HISQ solves"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 10x speedup is measured for the solve only, because the paper excludes eigensolve, disk I/O, and memory allocation from its timings, relying on those costs being amortized over many propagator solves per configuration; if a workflow computes only a few propagators per configuration, the gain may vanish or reverse.","fun_headline_variants_meta":{"raw":{"variants":["Deflation speeds up light-quark solves 10x on 144^3","33k CG iterations drop to 2.9k with deflation","Single-precision deflation halves memory without accuracy loss","Deflation: practical multigrid alternative for HISQ solves"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000838,"raw_usage":{"total_tokens":3633,"prompt_tokens":905,"completion_tokens":2728,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":2653}},"tokens_in":521,"tokens_out":2728,"duration_ms":22304,"temperature":1.0,"reasoning_tokens":2653,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T20:00:09.121226+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a production workflow at the physical light quark mass on the $144^3 \\times 288$ lattice that computes a realistic number of propagators per configuration, and record total wall-clock time including eigenvector generation, disk read, and memory allocation; if the deflated total is not below undeflated CG, the amortization premise fails. A specific crossover test is to find the number of solves per configuration at which the two totals are equal.","supporting_citations":[{"cited_title":"Numerical experiments using deflation with the HISQ action","cited_arxiv_id":"1710.07219","evidence_quote":"Earlier numerical experiments on deflation with the HISQ action; this work extends them to larger lattices and physical light quark masses."},{"cited_title":"Computing and deflating eigenvalues while solving multiple right hand side linear systems in Quantum Chromodynamics","cited_arxiv_id":"0707.0131","evidence_quote":"Develops the technique of computing and deflating eigenvalues while solving multiple right-hand side systems in QCD."},{"cited_title":"Optimizing Staggered Multigrid for Exascale performance","cited_arxiv_id":"2212.12559","evidence_quote":"Reports a 10x multigrid speedup for the same lattice size and quark mass, providing the comparison point for deflation versus multigrid."},{"cited_title":"Clark, R","cited_arxiv_id":null,"evidence_quote":"Describes the mixed-precision GPU solver machinery, including the reliable-update mechanism used in the CG baseline."},{"cited_title":"Stathopoulos, A.M","cited_arxiv_id":null,"evidence_quote":"Shows a multi-deflation approach for approximate eigenvectors that the paper adapts to single-precision eigenvectors."}],"review_version":1}