{"id":"45ad107a-d7aa-4da9-990f-64dece060c90","arxiv_id":"2412.19322","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey of mixed-precision numerical methods across CFD, climate, chemistry, and genomics, reporting speedups up to 8x on benchmarks and recommending co-design to unlock them.","lead":"This survey reviews how scientific computing can use low-precision and mixed-precision arithmetic to speed up simulation codes on modern AI-focused hardware. It catalogs domain results, algorithms, and software, and argues the gap between low and double precision performance is becoming too large to ignore.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central 8x speedup claim rests on the deliberately favorable HPL-MxP benchmark; the survey's own application-level evidence shows speedups of 1.1x–2x in most domains, so the claim of reshaping computational science is broader than the evidence supports.","rationale":"The paper is a survey, so the central claim is an interpretive synthesis rather than a new mathematical result. The most load-bearing part of that synthesis is the jump from benchmark-level speedups (8.31x on HPL-MxP) to the general promise that mixed precision can 'reshape computational science.' The survey itself provides strong internal evidence against the generality of this jump: Table 2 lists representative application speedups that are mostly in the 1.4x–1.9x range, and the text repeatedly notes that many applications are memory-bandwidth limited, where the theoretical ceiling is 2x for FP64-to-FP32. The HPL-MxP benchmark is explicitly designed to be compute-bound and to minimize the numerical difficulties of mixed precision, so it is a deliberately favorable case. The paper does not estimate what fraction of production scientific computing is compute-bound enough to approach such speedups, which leaves the central claim quantitatively ungrounded. This is a representativeness or framing problem, not a numerical error in the survey's reported facts. The reader's weakest assumption identified a related concern about benchmark speedups carrying over to production codes, but framed it jointly with the hardware-trend projection; my concern is more specifically that the survey's own data contradicts the broad applicability of the 8x headline, independent of future hardware evolution. The appropriate response is to condition the central claim on the minority of compute-bound, dense-linear-algebra-dominated workloads, and to either remove or clearly mark the private-communication performance numbers that the reader also flagged. None of this undermines the survey's genuine value as a structured review of the mixed-precision landscape; conditional acceptance with requests for these clarifications remains the right verdict.","tokens_in":44590,"tokens_out":4192,"duration_ms":37054,"concrete_test":"Build a workload-weighted speedup estimate from Table 2 and the bottleneck classifications in §2, using reported speedups and reasonable production workloads; compute both the arithmetic mean and an estimate weighted by compute-bound vs memory-bound fractions. If the aggregate falls below ~2x, the abstract should be revised to state that the 8x figure is specific to compute-bound dense linear algebra benchmarks, and the 'reshape' conclusion should be tempered. As a complementary check, re-run the HPL-MxP benchmark with a random dense matrix requiring partial pivoting and realistic condition number to measure how much of the 8.31x survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's headline '8x compared to double-precision in extreme compute-intensive workloads' is anchored in §3.2's HPL-MxP result of 8.31x on Frontier [71]. HPL-MxP is a strictly diagonal-dominant dense linear solve designed to eliminate pivoting and minimize iterative-refinement iterations, making it a best-case scenario. The survey's own §2.6/Table 2 reports representative application speedups of 1.44x (aerodynamics), 1.60x (other CFD), 1.40x (weather/climate), 1.9x (quantum chemistry) and 4.80x (genomics), and §1 concedes memory-bandwidth-limited applications cap at 2x for FP64-to-FP32. Thus the evidence in the paper itself indicates that an 8x-class speedup is not typical of the reviewed production scientific applications; it is an extreme upper bound. The central claim that mixed precision can broadly 'reshape computational science' would require either that a large fraction of scientific workloads are compute-bound dense linear algebra, or that memory-bound applications can be transformed to become compute-bound. The survey does not quantify that fraction, which is the load-bearing gap between the benchmark and the promise.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a survey of mixed-precision numerics in scientific computing, organized around four areas: application domains (CFD, weather and climate, quantum chemistry, genomics, AI), numerical algorithms (iterative refinement, splitting and emulation schemes, Krylov solvers, multigrid, FFT, eigensolvers, ODE solvers), resource utilization, and library support. It argues that the widening throughput gap between low-precision and FP64 hardware, together with recent algorithmic advances, makes mixed-precision methods an important opportunity for scientific computation. The abstract claims performance improvements of 8x over double precision in extreme compute-intensive workloads and concludes that mixed-precision numerics can reshape computational science.","tokens_in":44803,"tokens_out":6100,"duration_ms":55046,"significance":"The survey is timely, broad, and useful as a roadmap: it consolidates a large and scattered literature, proposes a clear taxonomy of mixed-precision algorithm classes (low precision, MxP-inL, MxP-R, splitting), and maps applications to computational motifs and libraries. The authors are also generally transparent about which figures are measured and which are estimated, for example by flagging assumptions in Table 2. The main weakness is that the headline 8x speedup claim is anchored in the deliberately favorable HPL-MxP benchmark, while the application-level evidence collected in the same manuscript shows speedups mostly in the 1.1x-2x range; this tension needs to be addressed before the survey's central promise can be accepted as stated.","major_comments":[{"comment":"The abstract's opening claim of '8x compared to double-precision in extreme compute-intensive workloads' is anchored in the HPL-MxP result of 8.31x on Frontier [71]. As the paper itself explains, HPL-MxP uses a strictly diagonal-dominant matrix, which removes pivoting and minimizes iterative-refinement iterations, and the survey's own Table 2 reports representative application speedups of 1.44x-4.80x while §1 notes that memory-bandwidth-limited applications cap at 2x when moving from FP64 to FP32. The 8x figure is therefore an extreme upper bound, not a typical outcome of the surveyed production codes. Since the concluding claim that mixed precision can 'reshape computational science' rests on broad applicability, the authors should either restrict the abstract to 'dense, compute-bound factorization workloads' or provide evidence on the fraction of scientific workloads that are compute-bound dense linear algebra.","section":"Abstract and §3.2"},{"comment":"Several load-bearing quantitative claims are attributed to private communications: the Blackwell HPL 1.8x performance-per-watt gain and the 4-34x GEMM energy-efficiency figures in §4, and the LSMS 'acceptable accuracy' statement in §2.3. In addition, the description of Ozaki II as 'ground-breaking' is the authors' reading of arXiv preprint [68], and the manuscript's own hedge ('if reproduced across different real-world matrices') indicates that this result has not yet been independently confirmed. For a survey that aims to guide adoption, these claims should be explicitly labeled as unverified and should not be used in support of the forward-looking hardware argument in §7 unless a public source is available.","section":"§3.1.2 and §4"},{"comment":"The asterisked speedups in Table 2 are labeled as 'reasonable assumptions by the current authors,' but the table is captioned 'Representative speedups obtained from mixed-precision methods.' The derivation of these entries is not transparent: for example, the 1.44x aerodynamics figure comes from multiplying the 1.2x speedup reported by Walden et al. by an assumed further 1.2x, as described in §2.1. Because the table is used to support the survey's assessment that application-level gains are modest, each assumed entry should either be replaced with a directly reported speedup or accompanied by a footnote showing the calculation, so that readers can distinguish measured results from the authors' estimates.","section":"Table 2 and §2.6"}],"minor_comments":[{"comment":"The 'Range' column entries such as '±10 ±308' appear to be missing superscripts; format as ±10^±308 and ±10^±38.","section":"Table 1"},{"comment":"The sentence ending in 'inverse problems.[]' contains an empty citation that should be filled or removed.","section":"§2.5"},{"comment":"The phrase 'preliminary results indicate that the results have acceptable accuracy [private communication]' should be marked as unpublished; as written, the bracket may be mistaken for a reference.","section":"§2.3"},{"comment":"The phrase 'they make a second-hand claim of speedups approaching 40%' should name the original source instead of describing it as second-hand.","section":"§2.2"},{"comment":"Equation (11) is referenced in §2.2 before it is defined in §3.8; add a forward reference or renumber the equations.","section":"§2.2 and §3.8"},{"comment":"The phrase 'these results should be taken with a pinch of salt' is too informal for a journal; rephrase as 'these results are preliminary and should be interpreted with caution.'","section":"§3.10"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript is publishable after revision. The main concern is not technical correctness but calibration of the claims against the evidence gathered in the same paper. The authors' own benchmark work is cited appropriately, but the abstract's 8x number and the prominence of private-communication results should be tempered. The survey is within scope for cs.CE and, once the overclaim is fixed, would be a useful reference for the community."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a well-organized survey of mixed-precision numerics in scientific computing. Its most genuinely useful contribution is the taxonomy in §3 (MxP-R, MxP-inL, split), which gives practitioners a clear vocabulary for thinking about where mixed precision can help. The coverage is broad—applications from CFD to genomics, algorithms from iterative refinement to Ozaki splitting, and a section on libraries—and it is the only survey I know of that pulls these threads together. The authors are also transparent about many assumptions, for example flagging which numbers in Table 2 are their own estimates.\n\nThe soft spot is the framing. The abstract claims an 8x performance improvement over double precision and says mixed precision can \"reshape computational science.\" That 8x figure comes from HPL-MxP, a benchmark deliberately constructed to be compute-bound and strictly diagonal-dominant. The survey's own Table 2 reports representative application speedups of 1.4x–1.9x for aerodynamics, CFD, weather/climate, and quantum chemistry, and 4.8x for compute-bound genomics. The paper also notes in §1 that memory-bandwidth-bound applications cap at about 2x for FP64-to-FP32. So the evidence in the paper itself says the 8x number is an extreme upper bound, not a typical outcome. The \"reshape computational science\" claim is stronger than the surveyed applications support.\n\nTwo smaller issues: several key performance numbers rest on private communications (NVIDIA Blackwell, LSMS), which a reader cannot verify; and the \"ground-breaking\" description of Ozaki Scheme II is too strong for a preprint that, as the authors themselves note, may not deliver FP64 accuracy for all matrices. The self-citations to HPL-MxP and HPG-MxP are fine—those are their own measured, reproducible results.\n\nNone of this sinks the survey. It is a solid entry point for computational scientists wondering whether mixed precision is worth pursuing, and the taxonomy alone is worth the price of admission. But the abstract and conclusions should be recalibrated to match the evidence: typical gains in real applications are modest, with large gains in specific compute-bound cases.\n\nI would send this to peer review. A good referee will ask the authors to tone down the general claim, replace private communications with public references or mark them clearly as unverified, and soften the Ozaki II language. After that, it is a publishable survey with real value for the HPC community.","headline":"A competent and useful survey of mixed-precision numerics, but the headline 8x speedup claim rests on a best-case benchmark and is not representative of the application-level evidence the survey itself reports.","tokens_in":45343,"tokens_out":1693,"would_cite":true,"duration_ms":16814,"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":"This survey argues that mixed-precision numerics can deliver up to 8x speedups in compute-intensive scientific workloads by aligning algorithms with hardware that now favors low-precision arithmetic.","keywords":["mixed precision","floating-point arithmetic","iterative refinement","tensor cores","HPL-MxP benchmark","scientific computing","GPU acceleration","numerical linear algebra"],"falsifier":"Measure the end-to-end speedup of a production scientific application that is known to be memory-bandwidth-limited when its solver is switched from double to single precision with iterative refinement. If the speedup stays at or below 2x while the hardware's FP16-to-FP64 throughput ratio is large, the claim that mixed-precision capabilities deliver 8x speedups in scientific workloads is falsified for that class of applications.","tokens_in":44366,"feed_emoji":"⚡","tokens_out":5663,"duration_ms":48705,"temperature":0.7,"pith_summary":"Modern high-performance computing hardware now devotes a large share of its silicon to low-precision arithmetic, driven by AI workloads, while most scientific simulations still run in double precision. This survey assembles evidence from fluid dynamics, weather and climate modeling, quantum chemistry, and genomics to argue that mixed-precision numerics can turn that hardware imbalance into speedups—up to $8\\times$ in extreme compute-intensive workloads—without losing the accuracy scientists need. The paper classifies existing techniques into three families: low precision used where the application tolerates it, low precision inside inner iterative components such as preconditioners, and splitting or refinement schemes that recover high-precision results from low-precision arithmetic. It concludes that compute-bound applications dominated by dense linear algebra are the nearest win, while memory-bound applications face a smaller ceiling of about $2\\times$ from halving the bit width.","feed_headline":"Mixed precision can speed science up to 8x","feed_subtitle":"Survey of CFD, climate, chemistry and genomics shows fast low-precision hardware can be used without losing accuracy.","key_machinery":"The workhorse is iterative refinement: solve a linear system in low precision, compute the residual in high precision, and correct the solution until the error reaches the desired level. Around that core, the survey organizes current methods into three classes: inner low precision, where approximate operations like preconditioners or multigrid smoothers run in low precision; mixed precision with refinement, where a high-precision correction step restores accuracy; and splitting schemes such as Ozaki splitting, which represent a high-precision matrix as a sum of several low-precision matrices and emulate high-precision matrix multiplication on tensor cores. The split count, and thus the speedup, depends on the dynamic range of the matrix entries, and the survey gives a threshold: emulation beats native double-precision GEMM only when low-precision throughput exceeds FP64 throughput by roughly 40–118x depending on the matrix and the splitting variant.","core_discovery":"The paper's central claim is that mixed-precision numerics can reshape computational science by aligning algorithms with the evolving hardware capability landscape. The authors support this by reviewing applications that have already adopted mixed-precision strategies and reporting the speedups they achieved, and by surveying algorithmic techniques—iterative refinement, splitting and emulation schemes, and adaptive precision solvers—that let a low-precision computation be corrected to full-precision accuracy. On a benchmark that isolates compute-bound dense factorization, the paper reports speedups of 9.50x and 8.31x over double-precision Linpack on the tested systems, and the survey argues that similar gains are available to production codes whose dominant motifs are dense matrix operations. For memory-bandwidth-limited applications, the paper itself sets the maximum speedup at 2x from double to single precision, since the gain comes only from moving fewer bits.","pith_inferences":["The paper's 8x figure comes from an idealized compute-bound benchmark; a careful reader should treat it as an upper bound, because most real scientific applications are at least partly memory-bound and therefore capped near the 2x bit-width ratio.","If hardware vendors continue shifting silicon toward low-precision tensor cores, the default scientific computing stack may eventually emulate FP64 arithmetic in software on low-precision units, making today's specialized splitting libraries a general-purpose foundation.","A testable extension of the survey's argument: instrument a production implicit CFD solver with adaptive-precision sparse matrix-vector products inside GMRES, and measure whether speedups in the 1.1x–6x range reported for benchmark matrices appear in the full application."],"forward_implications":["Compute-bound applications built on dense matrix multiplication can expect large speedups from mixed-precision LU factorization with iterative refinement, with reported values ranging from 3x to 9.5x on current GPU systems.","Memory-bandwidth-bound applications will see far smaller gains, at most about 2x from double to single precision, so the headline 8x claim does not extend to them.","As the low-precision to FP64 throughput ratio grows, emulating double precision with splitting schemes becomes competitive and eventually faster than native FP64 GEMM, as already demonstrated on one current platform.","Production scientific packages that have not yet adopted mixed-precision solvers, especially in CFD and quantum chemistry, can gain time-to-solution and energy savings by using existing libraries with mixed-precision iterative refinement.","The survey recommends co-design among domain scientists, numerical analysts, and computer scientists because the right precision choice is domain- and problem-specific."],"supporting_citations":[{"why":"Supplies the mathematical foundation of mixed-precision algorithms in numerical linear algebra, especially iterative refinement and precision formats.","marker":"[17]"},{"why":"Documents the hardware trend of a widening gap between low-precision and FP64 throughput that motivates the survey.","marker":"[9]"},{"why":"Provides the 2021 collection of mixed-precision algorithm advances across computational motifs from the exascale software effort.","marker":"[14]"},{"why":"Gives iterative refinement, the core correction mechanism underlying many mixed-precision solvers.","marker":"[61]"},{"why":"Supplies multiword arithmetic and the FABsum method for emulating high-precision GEMM on tensor cores.","marker":"[33]"},{"why":"Introduces the Ozaki splitting scheme, the central splitting technique for emulating high precision with low-precision matrix multiplications.","marker":"[63]"},{"why":"Demonstrates emulating DGEMM on integer tensor cores and quantifies the split count and performance thresholds.","marker":"[65]"},{"why":"Presents Ozaki Scheme II, achieving emulated GEMM faster than native double-precision GEMM on current hardware.","marker":"[68]"},{"why":"Reports the HPL-MxP benchmark speedups of 9.50x and 8.31x that anchor the survey's 8x claim.","marker":"[71]"}],"fun_headline_variants":["Mixed precision boosts science up to 8x with accuracy intact","Survey: mixed-precision numerics deliver up to 8x speedups","Faster science via mixed precision: survey spells out gains","Mixed precision: up to 8x faster computation, no accuracy loss","Mixed-precision algorithms reshape science with 8x speedups"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the expectation that the throughput gap between low-precision and double-precision arithmetic on future hardware will continue to widen, and that benchmark speedups like the 8.31x figure carry over to production scientific codes; the survey itself notes that many applications are memory-bandwidth-limited, where the maximum speedup from double to single precision is only 2x.","fun_headline_variants_meta":{"raw":{"variants":["Mixed precision boosts science up to 8x with accuracy intact","Survey: mixed-precision numerics deliver up to 8x speedups","Faster science via mixed precision: survey spells out gains","Mixed precision: up to 8x faster computation, no accuracy loss","Mixed-precision algorithms reshape science with 8x speedups"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000763,"raw_usage":{"total_tokens":3394,"prompt_tokens":965,"completion_tokens":2429,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":2339}},"tokens_in":581,"tokens_out":2429,"duration_ms":15336,"temperature":1.0,"reasoning_tokens":2339,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:41:05.141376+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the end-to-end speedup of a production scientific application that is known to be memory-bandwidth-limited when its solver is switched from double to single precision with iterative refinement. If the speedup stays at or below 2x while the hardware's FP16-to-FP64 throughput ratio is large, the claim that mixed-precision capabilities deliver 8x speedups in scientific workloads is falsified for that class of applications.","supporting_citations":[{"cited_title":"In: Proceed- ings of the International Conference on High Performance Computing, Networking, Storage and Analysis","cited_arxiv_id":null,"evidence_quote":"Reports the HPL-MxP benchmark speedups of 9.50x and 8.31x that anchor the survey's 8x claim."}],"review_version":1}