HSTU

2 articles tagged with this topic

NVIDIA's 5.93x AI Recommendation Speedup Only Holds With 100% Cache Hits

NVIDIA's HSTU recommendation stack peaks at 5.93x speedup — but only at 100% cache hit. Real-world hit rates hover near 50%. Don't pitch the peak.

Oct 12 min read

Recommenders Adopt the LLM Playbook — NVIDIA and Meta Ship HSTU as Turnkey

Recommenders shift to unified sequence modeling. NVIDIA and Meta ship Meta's 2024 HSTU as turnkey inference — feeds may enter an LLM-like era.

Sep 302 min read