# INT8 but no speedup: the time eaten by the memory wall on RK3588

- Source: Guyuehome
- Published: 2026-11-02T11:16:00.000Z
- RoboHot: https://robohot.news/en/items/rjvlitox0q33cfo4vnin13h4t
- Original: https://www.guyuehome.com/detail?id=2115988326080081922

## Summary

On RK3588, converting a model to INT8 does not speed up inference; the bottleneck comes from the memory wall rather than compute. The article analyzes the portion of time consumed by memory access in this phenomenon, explaining that quantization gains are limited by bandwidth rather than computational precision.