Gemma 12B runs much slower than expected in one local setup

Switching to Q8 cut generation speed from about 70 to about 10 . This happened even though the earlier model, GPT-OSS 20B Q4, was larger.

Trying a Q5 version of Gemma did not meaningfully improve speed compared with Q8. Turning off only added about 2 more .

The setup runs llama.cpp with 16 threads, an 8192 , 99 GPU layers, and a 4096 batch size. The GPU uses about 10GB of its available 20GB memory, so the slowdown does not look like a simple problem.

Key points

  • Q8 dropped from about 70 to 10 in this setup.
  • The Q5 version did not give a clear speed improvement over Q8.
  • Disabling only added about 2 .
  • The GPU used about 10GB out of 20GB, so memory capacity was not fully used.
  • Agent builders should benchmark real before switching models.
Read original