Gemma 4 MTP Slows Down on Dual 3090s
DjCanalex · reddit · 2026-07-19
The author tested the MTP (draft/multi-token prediction) performance of Gemma 4 31B QAT on dual and single RTX 3090 setups, concluding that it actually slows down in a dual-GPU environment.
Observed Results
- Dual 3090s: 39.7 t/s without MTP, dropping to a fluctuating 27–34 t/s with MTP enabled
- Single 3090: 33 t/s without MTP, initially jumping to 45 t/s with MTP enabled, then gradually settling at 36–37 t/s
Additional Logs
- Dual-GPU setup: draft acceptance rate of 0.56, mean length of 2.22
- Single-GPU setup: draft acceptance rate of 0.56, mean length of 2.29
The author shared output logs for both configurations along with their llama-server startup commands, seeking to understand why MTP performance on multiple GPUs contradicts the single-GPU results.
More from Infra
- Anthropic S-1: training infra edge cut per-task cost 40% and doubled revenue growth — BenBajarin · 2026-10-26
- Rewriting scaling law study: optimal synthetic data share rises from 10% to 32% with compute — heghbalz · 2026-10-26
- US states halt new data centers as 72% of Americans oppose local builds — YvesMulkers · 2026-10-26
- $5M Seed Round Can't Afford One DeepSeek Training Run: GPU-less AI Startup's Three Paths — FinanceYF5 · 2026-10-26
- HBM demand shifts: low-stack inference plans shelved, Rubin to go 12-high only — BenBajarin · 2026-10-26
- CircuitHub unveils The Grid, a factory-scale robot making PCBs in the US twice as fast as Shenzhen — garrytan · 2026-10-26