{ "schema_version": 1, "project": "K GraphFuse", "version": "0.1.0", "creator": "Utkarsh Sinha", "historical_benchmark_name": "Sinha GraphFuse", "evidence_url": "https://github.com/Utkarsh-Sinha0/k-graphfuse/tree/main/evidence", "claims": [ { "id": "same_t4_median", "kind": "measurement", "value": 21.281298617317614, "unit": "percent lower median latency", "source": "evidence/SINHA_GRAPHFUSE_PUBLIC_REFERENCE_RAW.json", "scope": "Original probability-score vectors; unweighted undirected global PageRank; uniform initialization; damping0.85; exactly10 iterations", "limitations": "One selected configuration on one graph in one run; no global SOTA or independent replication" }, { "id": "reference_pass", "kind": "measurement", "value": 11, "unit": "accepted configurations", "source": "evidence/SINHA_GRAPHFUSE_PUBLIC_REFERENCE_RAW.json", "limitations": "Reference acceptance does not imply universal accuracy improvement; aggregate L1 error is higher" }, { "id": "sampled_memory", "kind": "measurement", "value": 56.755862434095874, "unit": "percent lower sampled total GPU peak", "source": "evidence/SINHA_GRAPHFUSE_PUBLIC_REFERENCE_RAW.json", "limitations": "20ms sampling; inherited/library allocations included; short peaks may be missed" }, { "id": "broader_gate", "kind": "measurement", "value": 0, "unit": "passes among 19 complete strongest-control GPU gates", "source": "evidence/SINHA_GRAPHFUSE_COMPARISONS.csv", "limitations": "Different suite and gates than the matched public-reference task", "scope": "Historical 19-workload large/community/scale study; not the newer 15-suite study" }, { "id": "job_impact", "kind": "model", "value": 10.640649308658801, "unit": "percent modeled whole-job latency reduction at 50% matching-stage share", "source": "evidence/application_impact_model.csv", "limitations": "Amdahl projection, not measured customer savings" }, { "id": "complete_inequality", "kind": "conditional mathematical premise", "value": 2, "unit": "complete inequality constant", "source": "docs/concepts/mathematics.md", "limitations": "Depends on cited manuscript; public PageRank uses zero truncation and demonstrates no degree reduction" }, { "id": "native_speed_wins_20261010", "kind": "measurement", "value": 59, "denominator": 89, "comparable_cases": 85, "accuracy_failures": 4, "unit": "strong speed-gate wins", "source": "evidence/2026-10-10/comparison_summary.json", "scope": "Recorded resident ranking comparisons against accepted native cuGraph", "limitations": "59/85 comparable; four additional K accuracy failures counted as non-wins; not the fastest-control or full-workflow count" }, { "id": "native_speed_error_wins_20261010", "kind": "measurement", "value": 26, "denominator": 89, "unit": "speed and two-error wins", "source": "evidence/2026-10-10/native_comparisons.csv", "scope": "Strong speed gate plus recorded aggregate L1 and maximum-relative errors no worse; all 26 same precision", "limitations": "Rounded reference errors; no all-resource or downstream-quality claim" }, { "id": "strongest_overall_wins_20261010", "kind": "measurement", "value": 33, "denominator": 85, "unit": "strong speed-gate wins", "source": "evidence/2026-10-10/detailed/verification_summary.json", "scope": "Fastest accepted tested CPU/GPU control where measured", "limitations": "No exhaustive global SOTA comparison; CPU size limits apply" }, { "id": "strongest_gpu_wins_20261010", "kind": "measurement", "value": 38, "denominator": 85, "unit": "strong GPU speed-gate wins", "source": "evidence/2026-10-10/detailed/verification_summary.json", "scope": "Fastest accepted tested GPU control", "limitations": "Different count from native-only comparisons" }, { "id": "precision_dual_gates_20261010", "kind": "measurement", "value": 6, "denominator": 6, "unit": "speed and sampled-accuracy passes", "source": "experiments/precision_gpu/PRODUCT_VERIFIED.json", "scope": "Constructed 10k\u20131m-vertex FP64 polynomial filters and derivatives, 4/16 channels, Tesla T4", "limitations": "CuPy/PyTorch controls, not native cuGraph; first 16 rows use high-precision truth; later repaired API not fully retimed" }, { "id": "precision_speedup_range_20261010", "kind": "measurement", "value": [ 1.2497575244368624, 3.7394633833211692 ], "unit": "control median / selected median", "source": "experiments/precision_gpu/PRODUCT_VERIFIED.json", "scope": "Six selected T4 cases; full samples and p95 retained", "limitations": "Resident compute; validation, I/O, preparation and JIT excluded; profile policy reconstructed after exploration" }, { "id": "original_favorable_metrics", "kind": "measurement", "value": 8, "denominator": 9, "unit": "favorable observations", "source": "evidence/measured_metrics.csv", "scope": "Original matched ten-update T4 reference task; five latency, two sampled-memory and two error measures", "limitations": "Correlated measurements of one setup; aggregate L1 error worse, sampled power higher; later confirmation loses on median" }, { "id": "precision_time_reduction_range_20261010", "kind": "derived measurement", "value": [ 19.984478553101926, 73.2581951608292 ], "unit": "percent less median compute time", "source": "experiments/precision_gpu/PRODUCT_VERIFIED.json", "scope": "Six constructed Tesla T4 cases against tuned FP64 CuPy/PyTorch controls, not native cuGraph", "limitations": "Two conversions of the same isolated resident latency ratios; not observed concurrent throughput, end-to-end capacity or customer savings" }, { "id": "precision_derived_rate_range_20261010", "kind": "derived measurement", "value": [ 24.97575244368624, 273.9463383321169 ], "unit": "percent higher derived reciprocal-latency serial compute rate", "source": "experiments/precision_gpu/PRODUCT_VERIFIED.json", "scope": "Six constructed Tesla T4 cases against tuned FP64 CuPy/PyTorch controls, not native cuGraph", "limitations": "Two conversions of the same isolated resident latency ratios; not observed concurrent throughput, end-to-end capacity or customer savings" }, { "id": "controlled_packed_balanced_strong_speed_20261010", "kind": "measurement", "value": 18, "unit": "strong speed wins among 20 pooled device-fixture comparisons", "source": "studies/controlled_20261010/SUMMARY.json", "scope": "Matched-budget and matched-degree packed-balanced CuPy controls; ten fixtures from six families on two GPUs; not native cuGraph or exhaustive SOTA", "limitations": "Author-controlled two-device study; descriptive intervals, full FP64 checks and separate small arbitrary-precision checks; no universal accuracy, independent replication, cold-start or customer-cost claim" }, { "id": "controlled_packed_balanced_time_range_20261010", "kind": "measurement", "value": [ 36.88137721440219, 83.71266921830164 ], "unit": "percent less resident compute time in 18 winning comparisons", "source": "studies/controlled_20261010/SUMMARY.json", "scope": "Positive subset only; R-MAT uses 55.52% more time on RTX and 2.21% more on T4", "limitations": "Author-controlled two-device study; descriptive intervals, full FP64 checks and separate small arbitrary-precision checks; no universal accuracy, independent replication, cold-start or customer-cost claim" }, { "id": "controlled_application_time_reduction_20261010", "kind": "measurement", "value": [ -7.131392515329038, 0.07124116904921873 ], "unit": "percent lower pooled-median request time, RTX then T4", "source": "studies/controlled_20261010/SUMMARY.json", "scope": "Specialized scalar-restart CuPy comparison; RTX is slower, T4 is a practical tie; both intervals include one", "limitations": "Author-controlled two-device study; descriptive intervals, full FP64 checks and separate small arbitrary-precision checks; no universal accuracy, independent replication, cold-start or customer-cost claim" }, { "id": "controlled_application_top10_agreement_20261010", "kind": "measurement", "value": 1.0, "unit": "fraction exact agreement", "source": "studies/controlled_20261010/SUMMARY.json", "scope": "256 MovieLens 1M users, fixed finite diffusion; Recall@10=0.0587420862 and nDCG@10=0.0944984887; not recommendation quality SOTA", "limitations": "Author-controlled two-device study; descriptive intervals, full FP64 checks and separate small arbitrary-precision checks; no universal accuracy, independent replication, cold-start or customer-cost claim" }, { "id": "bound_attribution_time_range_20261011", "kind": "measurement", "value": [ 3.5604521480585904, 5.354788760780583 ], "unit": "percent less observed median resident time", "source": "studies/bound_attribution_20261011/SUMMARY.json", "scope": "Eight accepted matched FP64 packing/capture settings in two degree-changing T4 workloads; 31 samples/configuration; not a 10% strong-speed pass", "limitations": "Retrospective existing-data analysis; no independent repeat or paired-block interval; exact 9-to-8 GPU time absent and checked laptop old/new constants retain equal degrees. No new theorem authorship, algorithmic priority, native-cuGraph win or customer saving." }, { "id": "bound_attribution_matched_pairs_20261011", "kind": "measurement", "value": 8, "unit": "accepted matched timing pairs", "source": "studies/bound_attribution_20261011/SUMMARY.json", "scope": "Four settings in each of two constructed workloads; two of eleven examined records change degree; not eight independent graphs", "limitations": "Retrospective existing-data analysis; no independent repeat or paired-block interval; exact 9-to-8 GPU time absent and checked laptop old/new constants retain equal degrees. No new theorem authorship, algorithmic priority, native-cuGraph win or customer saving." }, { "id": "bound_attribution_degree_reductions_20261011", "kind": "measurement", "value": 2, "unit": "workload records with reduced degree", "source": "studies/bound_attribution_20261011/SUMMARY.json", "scope": "Eleven examined legacy matrix-filter/sensitivity records; nine unchanged; separate from the controlled study", "limitations": "Retrospective existing-data analysis; no independent repeat or paired-block interval; exact 9-to-8 GPU time absent and checked laptop old/new constants retain equal degrees. No new theorem authorship, algorithmic priority, native-cuGraph win or customer saving.", "denominator": 11 }, { "id": "bound_attribution_factor_reduction_20261011", "kind": "mathematical deduction", "value": 17.157287525380983, "unit": "percent smaller numerical-range value-tail factor", "source": "studies/bound_attribution_20261011/SUMMARY.json", "scope": "Fixed nonzero residual and geometry: 1+sqrt(2) to 2; ordinary norm branch may dominate; not a timing or execution-error percentage", "limitations": "Retrospective existing-data analysis; no independent repeat or paired-block interval; exact 9-to-8 GPU time absent and checked laptop old/new constants retain equal degrees. No new theorem authorship, algorithmic priority, native-cuGraph win or customer saving." } ], "unestablished": [ "global SOTA", "unique market positioning", "universal accuracy superiority", "integrated energy savings", "cloud-bill savings", "customer production quality", "guaranteed search or AI ranking" ], "evidence_updated": "2026-10-11" }