1st Workshop on Cross-Domain Architectural Support for Information Retrieval
Retrieval is a primitive. The same operation runs beneath attention, nearest-neighbor search, recommendation, databases, and tiered-memory access — this workshop treats it as one architectural problem.
Information retrieval is not merely an application — it is a fundamental computational primitive that recurs across modern workloads. Scaled dot-product attention in transformers is differentiable retrieval over a dynamic key–value store. Embedding-table lookup in recommendation systems is sparse retrieval under an inner-product metric. Approximate nearest-neighbor search underlies semantic search. Predicate evaluation in databases is selective retrieval over relational structure. Traditional CPU server-class workloads such as key-value stores and web-search engines retrieve records and documents from large collections, and even a remote access in tiered memory amounts to fetching a line or page from a far store.
A survey of recent MICRO, ISCA, HPCA, and ASPLOS papers reveals that many works — published in different tracks, framed around different applications — are, at a deep level, attacking the same family of primitive operations. RETRIEVE brings these threads together to ask: what is the minimal architectural substrate that efficiently supports all of them? Is there a convergent architecture, or a small and principled family, capable of accelerating retrieval across domains?
We welcome architecture and systems contributions on retrieval primitives, including but not limited to:
We solicit papers of up to six pages. Accepted papers appear in the workshop proceedings.
Novel contributions, preliminary or negative results, and research hypotheses or agendas are all welcome, across any retrieval workload — from attention and ANN search to recommendation, databases, key-value stores, web search, and tiered-memory access.
Submissions must use the two-column IEEE Computer Architecture Letters (CAL) format, and are reviewed double-blind.
Submissions are handled through HotCRP. We especially encourage submissions that connect insights across domains — for example, applying lessons from attention acceleration to ANN search hardware.