Ascend SuperPOD Optimization Boosts DeepSeek-V4 Training MFU to 34.22%

A full-stack optimization system named SLAI T-Rex successfully increased the training efficiency (MFU) of the trillion-parameter DeepSeek-V4 models to 34.22% on an Ascend NPU SuperPOD, effectively solving challenges in full-parameter post-training.

2026-07-23 ~ 2026-07-23 · 2 related posts

• DeepSeek-V4 post-training on Ascend SuperPOD reaches 34.22% MFU — Dongfang Li · 2026-07-23
• Ascend SuperPOD optimization lifts DeepSeek-V4 post-training MFU to 34.22% — pmttyji · 2026-07-23