AI Science Tracker
EN

An AI-written attention kernel accelerated prefill on an AMD GPU

University of California, Berkeley · 27 Oct 2026 · computer science

Discovery

An AI-written attention kernel accelerated prefill on an AMD GPU

Source check: Source checked. Not independent scientific confirmation.

Finding: A Claude Opus 5.5 team wrote a Kimi Delta Attention prefill kernel with fused input preparation and recurrent updates. Across 10 tested workloads on an AMD MI350X, it was 3.01 times as fast as the vLLM HIP kernel used for comparison.

Technology: Claude Opus 5.5; Evolving Agent Teams

Family Claude: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
Claude Opus 5.5 agents developed the fused implementation and tested its scheduled state handoffs, starting from a port of the vLLM baseline.
Test and result
The source reports five timing re-measurements and checks of outputs and final states under a numerical tolerance. The team added a slow-decay test after finding that earlier checks could miss state-handoff errors.
Limitation / still unknown
The reported speedup is an arithmetic mean for the tested hardware and workloads. The result concerns a single constructed kernel and tolerance-based validation, not universal throughput or exact equivalence.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.36848v1 · https://arxiv.org/html/2610.36848v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: §4.7 and Appendix C, MI350X construction, test definition and correctness gates.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open in the full catalog for filters, charts and related discoveries. Browse computer science discoveries.

Related catalog entries

Links use documented relationships where available, plus shared catalog context. A shared topic or model family does not establish a scientific dependency.