Chimera Boosts Multi-Vector Retrieval Throughput by 16x via GPU-CPU Co-Processing
_reachsumit · x · 2026-08-25
Chimera is a new multi-vector retrieval system designed to address bottlenecks in existing approaches through GPU-CPU co-processing. Unlike prior GPU-based systems like PLAID, which are limited by data transfer overhead during queries, Chimera stores highly compressed, low-precision quantization codes on the GPU while maintaining high-precision data in CPU memory.
By leveraging GPU-resident data for efficient candidate generation and filtering, and employing a collaborative scoring scheme, Chimera completely avoids vector data transfer and enables computation overlap. Experiments on real-world datasets demonstrate that Chimera achieves up to 16.0x higher Queries Per Second (QPS) than existing methods at the same recall level.
More from Infra
- Cheat sheet: VRAM requirements for different LLM context sizes — LeviTurk · 2026-08-25
- Meta's Data Center Uses Water for 800 Homes; Local Alfalfa Uses 400x More — Promptmethus · 2026-08-25
- Scaling Personal GPU Compute Amid Rising HBM Prices — Blues520 · 2026-08-25
- Smaller models could reshape deployment economics with high efficiency — eyishazyer · 2026-08-25
- Hugging Face libraries trade raw speed for broad compatibility and feature coverage — bclavie · 2026-08-25
- Fix 7900 XTX Linux Crashes via amdgpu.runpm=0 — Snoo_81913 · 2026-08-25