← back to the showcase

Sequence-search families: 5.77× on the measured CHR1 study

A fixed ten-variant sequence-search job took 763.5–765.1 seconds with stock Cas-OFFinder and 128.7–134.4 seconds with TAKT, including a fresh base calculation in every TAKT family. The complete reported hits matched under the predeclared full-row multiset contract below. These are measured results for one frozen workload and research build, rather than a general Cas-OFFinder replacement or a prediction for another search.

The measured workload

Complete family times

W is ten complete, unmodified stock command-line searches, one for each prepared variant. U includes a fresh full base calculation plus all ten variants, runtime preparation and validation, complete output delivery and cleanup. No base calculation is supplied free from an earlier run, and no cost is spread over requests that were not executed.

Interleaved pair W: ten stock searches, seconds U: fresh base plus ten variants, seconds W/U
1, W then U 764.024 128.747 5.93×
2, U then W 765.061 134.016 5.71×
3, W then U 763.489 134.429 5.68×

The primary factor is the sum of all three W times divided by the sum of all three U times: 2,292.574 / 397.192 = 5.77×. It is not the median of the three ratios, which is 5.71×. Each row measures the complete ten-variant family, rather than one guide or one edit.

Downloading data, preparing the frozen inputs, installation, compilation and independent output comparisons are outside both timing windows. The recorded software/runtime state was warmed for both paths; this is not a cold-install benchmark. U includes the sorting needed for its complete canonical output. W is the stock raw-output path; sorting its output for independent verification is outside W. These are finite observations on a shared host, not confidence bounds.

What matched

The declared contract is CANONICAL_EQUIVALENCE of the complete all-hit row multiset. It sorts full output lines in byte-lexicographic order with LC_ALL=C. It preserves every field and line occurrence: query, full accepted record/header, forward-leftmost coordinate, oriented target sequence and mismatch letter case, strand, mismatch count and guide identifier. Duplicate rows remain duplicated. There is no top-k selection, field removal, rounding or deduplication.

All 30 complete paired scenario comparisons across the three pairs passed: 60 complete stock/TAKT scenario outputs, plus three separately paid base outputs. The complete count vector for all 100 guide identifiers, including zero-hit entries, was also checked. Comparison uses the complete sorted files; hashes are additional records rather than substitutes for comparing the output.

Raw stock line order differed across the three repetitions of each scenario. The claim is equality of the full canonical row multiset, not raw file byte order. Canonicalisation was declared before the price measurements. Successful completion is required; diagnostic progress and elapsed-time messages on stderr are recorded separately and are outside the scientific output contract.

Measured scope

The 5.77× result applies only to this CHR1 input, these guide occurrences, ten fixed-length substitutions, search settings and pinned research build/runtime. This CHR1 study does not cover whole-genome jobs, 1,000 guides, other chromosomes or guide sets, another PAM, query length or mismatch limit, insertions/deletions, multiple FASTA records or files, GPU operation, other runtime builds or general command-line compatibility. It includes no single-edit result, crossover family size or separate qualified phase or harness-share claim.

The comparison establishes the runtime and output contract of this finite study. Biological guide quality was not assessed. Another workload needs its own complete comparison under an agreed output contract.

Second reference: 6.07× on a CHR2 study with a re-pinned research build

A separate Cas-OFFinder 2.4.1 study measured a ten-variant family on GRCh38/hg38 chromosome 2: one FASTA record, chr2, with 242,193,529 bases. It used the same first 100 preselected guide occurrences as the CHR1 study, in the same order and with the same identifiers. Selection did not depend on CHR2 hits or timings; this does not assert that all 100 sequences are unique.

The search settings remained 20-base guides with an NGG PAM, the 23-position pattern 21 × N + GG, and at most four mismatches (L23/m4), including both strands and every reported hit. The ten inputs were predeclared, independent fixed-length substitutions of the same original CHR2 reference. They were not applied cumulatively.

This was a separately pinned research build. Only the reference-length and reference-header bindings and the build epoch changed. The source, compiler and runtime recipe otherwise remained the same. It was a new binary, rather than the CHR1 executable used unchanged. The study retained the pinned Linux x86-64 CPU OpenCL runtime and one assigned CPU; it includes no GPU comparison.

Complete family times

W is ten complete, unmodified stock command-line searches, one for each prepared variant. U includes a fresh complete base calculation and all ten variants, runtime preparation and checks, copies, sorting for canonical output, complete output delivery and cleanup. Each U pays for its own base calculation; no computed base is supplied free from a control or a previous family.

Pair Order W: ten stock searches, seconds U: fresh base plus ten variants, seconds W/U
1 W then U 764.8 126.4 6.051×
2 U then W 763.3 125.5 6.082×
3 W then U 762.8 125.6 6.075×

Times in the table are rounded to 0.1 seconds. Ratios use the unrounded times. The primary factor is the sum of all three W times divided by the sum of all three U times: summed times were 2,290.951 seconds and 377.456 seconds, giving 6.069455935×, or 6.07×. All three pairs are included. These are finite observations on a shared host, not confidence bounds or a predicted factor for another job. The separate CHR1 result remains 5.77×; the two studies are not pooled.

Downloading and preparing the frozen references and variants, installation, compilation and independent output comparisons are outside both timing windows. The recorded software/runtime state was warmed for both paths; this is not a cold-install benchmark. U pays for its full canonical sorting. W is the stock raw-output path, with sorting for independent verification outside W.

Complete output and controls

The contract remains CANONICAL_EQUIVALENCE of the complete all-hit row multiset: full output lines sorted in byte-lexicographic order with LC_ALL=C. Every field and row occurrence is preserved, including query, record/header, forward-leftmost coordinate, oriented target and mismatch letter case, strand, mismatch count and guide identifier. Duplicate rows remain duplicated. Raw line order is not the contract; there is no field removal, rounding, top-k selection or deduplication.

All 30 complete paired variant comparisons passed, covering 60 stock/research variant outputs. The complete 100-entry guide-occurrence count vector, including zero-hit entries, was also checked for each output. Full files were compared; hashes are additional records, not a replacement for that comparison. All three freshly calculated, paid base outputs also matched an independently generated stock reference.

A separate edited-input control matched fresh stock output. A forced late-refusal control completed through fallback and matched the complete stock result under the same contract. There were zero fallbacks in the three priced families. Diagnostic progress and elapsed-time messages on stderr are outside the scientific-output contract.

Scope of this second study

The 6.07× factor covers this CHR2 reference, these guide occurrences, the ten fixed-length edits, the stated search settings and the newly pinned research build/runtime. It does not establish a result for other references or guide sets, whole-genome searches, 1,000 guides, other PAMs, query lengths or mismatch limits, insertions/deletions, multiple FASTA records or files, GPU execution or another runtime. Single-edit performance and a crossover family size were not measured. No qualified campaign-specific phase-share or harness-overhead estimate is claimed.

This is a fixed-workload research result, not a general compatibility release or a promise for another workload. Biological guide quality was not assessed. Another workload needs its own complete comparison under an agreed output contract. The existing software and reference-provenance notices on this page also apply to this study; this addition distributes no engine, reference, guide data or accelerator binary.

CHR2 measurement record fingerprints

These SHA-256 values identify the retained internal records. The records themselves are not included on this page.

Record SHA-256
CHR2 three-pair report, BUILD_RU.md 00c27d5786882e7d4670e9ea3912c9ba69205db798f6baf68aa2017540fdd27a
CHR2 final unit seal, SEAL.json b3c3bdc99f52015dcc38b39898d89a44b63b2dc32e0984347a3691b454e1ecf6
Independent source, contract and price review, REVIEW_CAS_OFFINDER_CHR2_T20_RU.md ca90d14b97d71bf3960b5c1101ed6469b3487b1642690a2aa6b4656b0a5e2701

Software and reference provenance

Cas-OFFinder 2.4.1 uses the new BSD / BSD-3-Clause licence. Its upstream notice has a separate exception for dirent.h; other components retain their own licence notices. The UCSC hg38 chromosome directory identifies the public reference source and its public-use notice.

This page contains result text and record fingerprints. It supplies no engine or TAKT binary, reference sequence files or guide data. Project names identify the measured software and data sources; no endorsement is claimed.

Measurement record fingerprints

These SHA-256 values identify the retained internal measurement records. The records themselves are not included on this page.

Record SHA-256
Three-pair measurement report 03db91799d14e4ec76e1b5fdef267018e8f6fd6d492bcc8f8d858a05675ccb09
Complete-family result metadata cfb29db4266aabf2ed01035a69ceb25915ebf2c7eca44e93a51110e419746d5e
Independent source and price review of this fixed study e94be0f71254ba0632257d5ed14a13616822b34d7b8a43745b6a03fb07a3503c

Describe a sequence-search workload: engine and version, reference and query sizes, variant family, required full output, runtime and CPU allocation. The recorded factors do not promise a gain for another workload.

Source text: README.md

Telegram