---
title: "Does a primary source confirm CRISP's '20 hours to less than a second' DNA-alignment speedup from processing-in-memory?"
type: "question"
status: "answered"
date_raised: "2026-07-11T00:00:00.000Z"
writer_model: "claude-opus-4-8"
tags: ["processing-in-memory","genomics","sequence-alignment","verification","quant"]
answered_log: [{"date":"2026-07-30T00:00:00.000Z","by":"[[claim-crisp-20-hours-to-under-a-second-figure-has-no-primary-source]], [[claim-rapid-rapidx-speedups-fall-short-of-crisp-press-figure]]","settled_by":"A direct search for the primary behind the UVA Today figure, and a check of CRISP's own peer-reviewed PIM sequence-alignment output (RAPID, RAPIDx), found no paper stating or reproducing '20 hours to less than a second.' The closest Tier-1 analogue — RAPID's own measured chromosome-wide alignment runtime (1081s, ~18 min) and its largest reported speedup ratio (~300×) — is roughly two to three orders of magnitude short of what the press figure implies. Answered in the negative: no primary confirms the figure, and the nearest available primary evidence argues against its magnitude."}]
---


The capture [[10-inbox/raw/2026-07-09-hop-memory-wall-pim-genomics]] surfaced a striking figure from UVA Today (Tier 3, university press): the CRISP center (Center for Research in Intelligent Storage and Processing in Memory) claims processing-in-memory redesigns cut DNA sequence-alignment time "from 20 hours to less than a second." This is a quantitative claim resting on a Tier-3 secondary source, so it was flagged `[unverified-quant — needs primary]` at capture and deliberately **not** promoted to a claim-note.

Specific verification needed:

- **Locate the primary** behind the UVA press figure — which CRISP publication, benchmark, or dataset produces the "20 hours → <1 second" number, and for which alignment algorithm and problem size? The press release does not cite the underlying paper.
- **Confirm the comparison baseline.** 20 hours on *what* hardware, against *what* PIM configuration? A ~72,000× wall-clock reduction is far larger than the "up to 4.06× / up to 28.14×" figures in the peer-reviewed [[claim-processing-in-memory-beats-memory-wall-dna-alignment|Bioinformatics PIM study]] — the two are likely measuring different things (end-to-end pipeline vs. a single alignment kernel; different baselines). Reconcile them.
- **Check whether it is a projection or a measurement.** University press figures on in-house hardware are frequently best-case or simulated.

Why it matters: the reconciled figure would either corroborate or badly undercut the cross-domain memory-wall bridge. Until a primary lands, the 4.06×/28.14× Tier-1 numbers are the only defensible quant on PIM sequence alignment; the 20hr→1sec figure stays out of the note graph.

**2026-07-30 — answered.** A follow-up capture ([[10-inbox/raw/2026-07-30-does-a-primary-source-confirm-crisps-20-hours]]) searched directly for the primary and checked CRISP's own peer-reviewed PIM alignment output (RAPID, ISLPED 2019; RAPIDx, IEEE TCAD 2023). Neither paper — nor Sieve, a third CRISP-affiliated paper on an adjacent workload — contains the figure. RAPID's own measured chromosome-wide alignment runtime (1081s, ~18 min on one chip) and its largest reported speedup ratio (~300× over CUDAlign 4.0 with 48 GPUs) are the closest Tier-1 analogue located, and both are roughly two to three orders of magnitude short of what "20 hours to under a second" implies. Promoted to [[claim-crisp-20-hours-to-under-a-second-figure-has-no-primary-source]] and [[claim-rapid-rapidx-speedups-fall-short-of-crisp-press-figure]]. The comparison-baseline and projection-vs-measurement sub-questions are moot given no primary was found to have either property.


## Progress log

- [object Object]