Pinterest (Feb 2026): Serving Two-Tower Models Using GPUs
> Source: https://lnkd.in/g4eyypSh
Overview
This is a stub/index entry for a Pinterest engineering resource dated February 2026, titled "Serving two tower models using GPU." The original linked resource describes how Pinterest serves two-tower models on GPU infrastructure. As archived on this forum, the entry contains only the title and link — no abstract, figures, or experimental results are included.
Background: Two-Tower Models
Two-tower (dual-encoder) architectures are widely used in large-scale retrieval and recommendation:
- A user tower encodes user context/history into an embedding.
- An item tower encodes candidate items into the same vector space.
- Item embeddings can be precomputed and indexed for approximate nearest-neighbor (ANN) retrieval, enabling millisecond-level candidate generation.
- The archived text contains no original content beyond the title; the linked source should be consulted for actual technical details.
- Any quantitative conclusions in this entry are not supported by the archive and should be verified against the original publication.
- Foundation Model for Personalized Recommendation by Netflix (Mar 2025)
- Evaluating Search Relevance Part 2 — Phi-3 as Relevance Judge
- Investigating ChatGPT Search: Insights from 80 Million Clickstream Records
Serving such models on GPUs raises engineering questions around embedding refresh frequency, index versioning, p99 latency budgets, and cost per query.
Caveats
Related Entries
Glossary
| Term | Meaning | |------|---------| | Two-tower model | Dual-encoder architecture with separate user and item encoders | | ANN | Approximate Nearest Neighbor search | | IR | Information Retrieval | | RAG | Retrieval-Augmented Generation |