Meta Paper: Feature-Importance Embedding Pruning Cuts Recommendation Model Memory 20-40%
_reachsumit · x · 2026-11-20
Meta researchers published FI-ITEP, an in-training embedding pruning method that uses feature importance (instead of size-based heuristics) to set per-table pruning ratios in large recommendation models.
- 20–40% total memory reduction across three production-scale models, with equal or better accuracy vs. percentile-based ITEP
- The FI signal is cheap to compute, so savings translate to systems gains: training dropped from 112 to 80 GPU trainers and saves 1,000+ GPU-hours per model launch
- Trends reproduced on open-source DLRM (Criteo 1TB), with end-to-end training-time and GPU-cost tradeoffs quantified
More from Infra
- Quantum Computing Takes Aim at AI Workloads as IBM and QuEra Map Fault-Tolerant Roadmaps — mikeflache · 2026-11-20
- Tencent Earnings Call: Latest Chip Restrictions Will Accelerate China's Domestic Semiconductor Supply Chain — basedjensen · 2026-11-20
- Chorus paper: shared n-gram tables cut memory 87% while beating DeepSeek's Engram — brown2green · 2026-11-20
- Tsinghua-spinoff Qingyi launches Ascend 950DT 10,000-GPU token factory, lifting utilization from 40% to 80% — 量子位 · 2026-11-20
- SenseNova-Loop 9B runs on 8GB VRAM at 22 t/s, uses half the tokens of Qwen3.8 — R_Duncan · 2026-11-20
- DeepSeek job posting reveals 100k-accelerator cluster challenges amid Ascen 950 training rumor — teortaxesTex · 2026-11-20