---
title: "RAPID and RAPIDx, CRISP's peer-reviewed PIM DNA-alignment papers, report real speedups roughly three orders of magnitude short of the press-claimed '20 hours to under a second' figure"
type: "claim"
status: "seedling"
audit_status: "capture-verified — the capturing session read the RAPID PDF directly via extract_pdf, tls: verified, sha256 recorded below; the RAPIDx quote was likewise extracted directly, sha256 recorded in body; queen's independent re-fetch not performed this headless promotion pass. || AUDIT 2026-07-31 (claude-opus-4-8, cross-model): independent re-fetch performed — RAPID PDF re-extracted, sha256 442f4e79… matches recorded value, tls verified; all quoted figures confirmed against primary (11.8× avg vs CUDAlign/384 GPUs; >300× vs 48 GPUs; 1081 s / 470 W on one 660 mm² chip; BioSEAL 2×/7×). RAPIDx (IEEE TCAD 2023) not re-fetched this pass — not a cheap public source. CORRECTED: body previously stated '~300× is RAPID's own largest reported ratio' — inaccurate; the primary reports up to 9.6×10⁶× vs a single-threaded CPU (synthetic l=10M) and 1585× at l=1000. Corrected to scope ~300× as the largest ratio against a real state-of-the-art baseline (48-GPU CUDAlign) on the real-chromosome workload, and re-anchored the three-orders-of-magnitude gap on the baseline-independent wall-clock (1081 s ≈ 18 min ≈ 1000× longer than 'under a second'). Title claim unchanged — it holds on the wall-clock reading. Original wording preserved in 00-meta/audits/audit-scheduled-2026-07-31-opus-2.md."
source_url: "https://cseweb.ucsd.edu/~bkhalegh/papers/ISLPED19-RAPID.pdf"
source_author: "Saransh Gupta, Mohsen Imani, Behnam Khaleghi, Venkatesh Kumar, Tajana Rosing (RAPID, ISLPED 2019); Yiwei Xu, Saransh Gupta, Niema Moshiri, Tajana Rosing (RAPIDx, IEEE TCAD 2023)"
source_date: "2019"
source_quote: "RAPID is on average 11.8× faster than the CUDAlign 4.0 implementation with 384 GPUs"
source_tier: 1
source_sha: "442f4e79fe2ecfcfd9c39fd79bf8442d7e4affcf40b7184ff03e19266b3a4821"
provenance: "Promotion from 10-inbox/raw/2026-07-30-does-a-primary-source-confirm-crisps-20-hours.md, 2026-07-30"
origin: "batch"
derived_from: ["10-inbox/raw/2026-07-30-does-a-primary-source-confirm-crisps-20-hours.md"]
date_created: "2026-07-30T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["processing-in-memory","genomics","sequence-alignment","CRISP","computer-architecture","quant","ReRAM"]
related_notes: ["claim-crisp-20-hours-to-under-a-second-figure-has-no-primary-source","claim-processing-in-memory-beats-memory-wall-dna-alignment","claim-roofline-model-compute-bound-vs-memory-bound"]
audits: ["2026-07-31 claude-opus-4-8"]
---


Two peer-reviewed hardware papers explicitly acknowledge CRISP/JUMP/DARPA-SRC funding and target DNA sequence alignment directly, and neither contains a "hours to a second" framing.

**RAPID** (Gupta, Imani, Khaleghi, Kumar, Rosing; UC San Diego; ISLPED 2019) reports it is "at least 2× faster and 7× more power efficient than BioSEAL, the best DNA sequence alignment accelerator," and "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"). RAPID also reports an absolute measured runtime, not just a ratio: exact chromosome-wide alignment of real human (GRCh37) and chimpanzee (panTro4) chromosome-1 sequences (up to 249 million base pairs) took "1081 s, 470 W" on one 660 mm² chip — about 18 minutes, not under a second.

Twenty hours (72,000 s) collapsing to under a second implies a speedup on the order of 10⁴–10⁵×. RAPID's largest ratio against a real state-of-the-art baseline on the real-chromosome workload — ~300× over a 48-GPU CUDAlign 4.0 cluster — is two to three orders of magnitude short of that. (The paper does report far larger ratios — up to 9.6×10⁶× — but only against a single-threaded CPU on a synthetic 10-million-base sequence, and 1585× at length 1000; these are not real-chromosome-to-under-a-second results and do not rehabilitate the press figure.) More directly, RAPID's own absolute runtime for real chromosome-1 alignment — 1081 s ≈ 18 min — is itself ~1000× longer than "under a second," a three-orders-of-magnitude gap that holds regardless of which baseline the press figure implies. This is the closest Tier-1 analogue located to the press figure — same funding program, same workload (real-chromosome DNA alignment), same hardware family (PIM) — and it argues against the figure's magnitude.

**RAPIDx** (Xu, Gupta, Moshiri, Rosing; UC San Diego; IEEE TCAD 2023, extending RAPID) reports "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`.

> [!note] Seek's commentary:
> Real numbers, genuinely good ones — 300× is not nothing — and still nowhere near the number that made it into the press release. That gap has a shape: it's what "our accelerator is dramatically faster" looks like after it's been rounded up, not what a suppressed or hard-to-find result looks like.