Does a primary source confirm CRISP's '20 hours to less than a second' DNA-alignment speedup from processing-in-memory?

This capture resolves the open item in 50-questions/question-verify-crisp-dna-alignment-20hr-to-1sec, itself routed from 10-inbox/raw/2026-07-09-hop-memory-wall-pim-genomics. It searched directly for the primary paper behind the UVA press figure and, failing that, checked whether CRISP's own peer-reviewed processing-in-memory (PIM) publications on DNA sequence alignment corroborate it.

Bottom line up front: no primary source stating or reproducing "20 hours to less than a second" was located. The figure traces to a single Tier-3 university press article with no cited study. CRISP's own Tier-1, peer-reviewed PIM sequence-alignment papers — the natural place such a result would appear — report real but far smaller and differently-shaped speedups, and the one measurement most directly comparable (full chromosome-wide alignment, real hardware) lands roughly three orders of magnitude away from "less than a second." The central question is marked [unverified — could not confirm; the closest Tier-1 evidence available cuts against the figure's specific magnitude].

Claim: The "20 hours to less than a second" figure traces to a single Tier-3 UVA press article that cites no underlying study, paper, or dataset

UVA Today (Audra Book, published 2022-02-08, "Say Goodbye to the Memory Wall") states that CRISP center researchers' PIM redesigns "could shorten sequence alignment time from 20 hours to less than a second," framed as a measured/"stunning" result rather than a projection. No paper, benchmark name, or dataset is cited alongside the figure, and Kevin Skadron (CRISP's director, quoted elsewhere in the piece) is not quoted attributing the number directly.

This is a quantitative claim resting on a Tier-3 university-press source: [unverified-quant — needs primary]. Additionally, the exact phrase above was retrieved this session via WebFetch (a fetch-and-summarize tool), not archive_page — the mcp__seek__archive_page tool returned a permissions error on every attempt against this URL this session (not a content-based safety flag; see Safety flags below). Per the quote-provenance rule, a quote obtained through a summarizing layer is not admissible as a sourced source_quote, so the phrase above is recorded as [unverified-quote — needs direct read] and is not carried in this note's frontmatter as source_quote. (A near-identical quote was captured by an earlier Seek session on 2026-07-09, predating the 2026-07-20 receipts rule — see 10-inbox/raw/2026-07-09-hop-memory-wall-pim-genomics.md — but that earlier fetch also cannot supply a sha256 receipt.)

Claim: CRISP's own Tier-1, peer-reviewed PIM sequence-alignment papers do not contain the "20 hours to less than a second" figure

Two peer-reviewed hardware papers were located that explicitly acknowledge CRISP/JUMP/DARPA-SRC funding and target DNA sequence alignment specifically:

• RAPID (Gupta, Imani, Khaleghi, Kumar, Rosing; UC San Diego; ISLPED 2019): "This work was partially supported by CRISP, one of six centers in JUMP, an SRC program sponsored by DARPA, and also NSF grants #1730158 and #1527034." Reported results: "RAPID is at least 2× faster and 7× more power efficient than BioSEAL, the best DNA sequence alignment accelerator," and "RAPID is on average 11.8× faster than the CUDAlign 4.0 implementation with 384 GPUs" (up to "over 300× faster than CUDAlign 4.0 with 48 GPUs"). No "hours to a second" framing appears anywhere in the paper. — Tier 1, direct quotes from extract_pdf, sha256 442f4e79fe2ecfcfd9c39fd79bf8442d7e4affcf40b7184ff03e19266b3a4821.
• RAPIDx (Xu, Gupta, Moshiri, Rosing; UC San Diego; IEEE TCAD 2023, extending the ISLPED 2019 paper): co-author Tajana Rosing's bio states she is "leading a number of projects, including efforts funded by DARPA/SRC JUMP CRISP program with focus on design of accelerators for analysis of big data." Reported results: "RAPIDx delivers 131.1× and 46.8× throughput improvements over state-of-the-art CPU and GPU libraries" for short-read alignment, "1.8-2.9× higher" than ASIC accelerators for long-read alignment, and "up to 321× speedup over Edlib" for edit distance — again, no "hours to a second" figure. — Tier 1, direct quotes from extract_pdf, sha256 de4fcc3ba30aa717b631fc8a99924e9b1fc5bd0012f1c76eeef7bdc959dd3f04.

A third CRISP-affiliated paper, Sieve (Wu, Sharifi, Lenjani, Skadron, Venkat; University of Virginia; ISCA 2021, with Skadron — CRISP's director — as co-author), targets a related but distinct workload (in-DRAM k-mer matching for metagenomic classification, not pairwise sequence alignment) and reports "an average of 326×/32× speedup... over multi-core-CPU/GPU baselines" — again a different figure, on a different task. — Tier 1, sha256 00be5d2e2af60f93c0fd00993e9c24b715310cf27bc4f9c7c526864992f1d52c. (Kept here as context rather than a core claim, since it is not the alignment workload the press figure describes.)

Claim: The closest Tier-1 analogue to the press figure — RAPID's own measured chromosome-wide alignment time — is roughly three orders of magnitude slower than "less than a second"

RAPID's paper reports an absolute measured runtime, not just a ratio: "One 660 mm² RAPID chip: 1081 s, 470 W" for exact chromosome-wide alignment of real human (GRCh37) and chimpanzee (panTro4) chromosome-1 sequences (up to 249 million base pairs). 1081 seconds is about 18 minutes — not "less than a second." Twenty hours (72,000 seconds) divided by "less than a second" implies a speedup on the rough order of 10⁴–10⁵×; RAPID's own largest reported ratio in the paper is "over 300× faster than CUDAlign 4.0 with 48 GPUs," roughly two to three orders of magnitude short of what the press figure implies. This is the single most directly comparable Tier-1 measurement located (same institution's funding program, same workload — DNA sequence alignment on real chromosome data, same hardware family — PIM), and it does not support the press figure's magnitude. — Tier 1, quotes and arithmetic as above, sha256 442f4e79fe2ecfcfd9c39fd79bf8442d7e4affcf40b7184ff03e19266b3a4821.

This also sits alongside the vault's existing claim-processing-in-memory-beats-memory-wall-dna-alignment (Diab et al., Bioinformatics 2023, UPMEM DIMMs, unrelated to CRISP): "up to 4.06× speedup" including data-transfer overhead, "up to 28.14× speedup" for alignment compute alone — again nowhere near a 10⁴–10⁵× figure. No Tier-1 measured PIM genomics benchmark found in this search or the prior one comes close to the multiplier the UVA press figure implies.

Further leads

• "Five Years that Changed Computing" (UVA Engineering, on CRISP's 2023 conclusion) repeats a version of the claim and adds "further speedups of 100 times or more believed to be possible" — worth a direct fetch to see if this softer, explicitly-projected language is the more accurate framing of the underlying result. WebFetch on this URL returned HTTP 403 this session; needs a different retrieval route. — https://engineering.virginia.edu/news-events/news/five-years-changed-computing
• CRISP's own 2018 GOMACTech program overview (Skadron, Xie, Martínez, Swanson, Patel — the center's founding PIs) names genomics/precision medicine as a target application area but predates any results; not itself evidence for or against the figure. — http://people.ece.cornell.edu/martinez/doc/gomactech18.pdf
• "Accelerating DNA Read Mapping with Digital Processing-in-Memory" (arXiv 2411.03832, 2024) is a more recent PIM DNA read-mapping paper, fetched but not checked against this question — worth a pass to see if it cites the CRISP figure or supersedes RAPID/RAPIDx's numbers. — https://arxiv.org/pdf/2411.03832
• BIMSA, GeneF, PIM-Align, and other 2020–2025 PIM sequence-alignment accelerators surfaced during search but were not checked for CRISP funding acknowledgments or the specific figure.

Safety flags

None. The only anomaly encountered was tooling, not content: mcp__seek__archive_page returned a permissions error on every attempt this session (against news.virginia.edu and engineering.virginia.edu URLs), and a plain WebFetch against the "Five Years that Changed Computing" URL returned HTTP 403. Neither is a recognition-rule signal per the safety spec — no addressed-to-AI language, override language, claimed authority, or urgency framing was observed on any page read this session. Noted here only because it affects what could be quoted-with-receipt.

Entity candidates

• CUDAlign 4.0 (de Oliveira Sandes et al.) — concept/system — the GPU-cluster baseline RAPID's and RAPIDx's speedup claims are measured against; this is the foundational prior-art system the priority/speedup comparison actually rests on, not just a footnote citation.
• BioSEAL / PRINS (Kaplan et al.) — concept/system — the ReRAM associative-computing accelerators RAPID explicitly benchmarks itself against as "the best DNA sequence alignment accelerator" prior to RAPID.
• CRISP (Center for Research in Intelligent Storage and Processing in Memory) — concept/org — the DARPA/SRC JUMP-funded, 9-university, $29.7M center (2018–2023) at the center of this whole question; already referenced but not yet a standalone entity page.
• Kevin Skadron — person — CRISP's director (UVA), co-author on Sieve, quoted (but not on the disputed figure) in the UVA Today piece.
• Tajana Rosing (Šimunić Rosing) — person — UCSD PI leading CRISP-funded PIM accelerator work; senior author on both RAPID and RAPIDx.
• Samira Khan — person — UVA CRISP PI named in press coverage as working on genomics; not yet confirmed as author on a specific alignment paper in this search.
• Audra Book — person — UVA Today writer, byline on the source article carrying the disputed figure.
• RAPID / RAPIDx — concept — the CRISP-funded PIM DNA-alignment architecture family this capture treats as the closest primary analogue.
• Sieve — concept — CRISP-affiliated (Skadron co-author) in-DRAM k-mer-matching accelerator; distinct workload from pairwise alignment, flagged as a further lead.