Combined, these techniques raised on‑prem throughput to , i.e., ≈ 95 % of the FFmpeg baseline, while reducing the CPU hot‑spot ratio to ≈ 28 % .
