Running 12 agents on one RTX 3090 lifts total throughput from 30 to 110 tok/s
Mchanger · reddit · 2026-10-31
On a single RTX 3090 (24GB VRAM) running Qwen3 27B Q4KM with 197k context, the author measured that running multiple agents concurrently yields much higher aggregate token throughput than a single agent, even though each agent gets lower tps and a smaller cache.
Measured results:
| Agents | Total tok/s | Per agent |
|---|---|---|
| 1 | 30 | 30 |
| 2 | 46 | 23 |
| 4 | 64 | 16 |
| 6 | 74 | 12 |
| 8 | 78 | 10 |
| 12 | 110 | 9 |
| 16 | 111 | 7 |
The author notes more agents means less context each, so efficient auto-compaction/looping is required. From there, swarms can let each agent spawn sub-agents for small tasks up to the machine's optimal agent count (12 here), maximizing hardware burn-rate without extra energy.
More from coding & agent
- Factory's Missions: The Emerging Primitive for Decomposing Large-Scale Agent Work — blaizedsouza · 2026-10-31
- Learner bot mines agent traces for recurring fixes and turns them into one-tap skill updates — bfrench · 2026-10-31
- Testing whether T3 code beats ocra — and the CLI coding tools' mobile UX problem — lxfater · 2026-10-31
- PentestCode Review: A Terminal AI Agent for Pentesting and Bug Bounty — glenbeer · 2026-10-31
- Unanswerable questions inflate Recall@K: dev open-sources RAG evaluator to fix it — infrixo · 2026-10-31
- Yacine's advice to new programmers: push hardware to the limit, bitbang GPIO — yacineMTB · 2026-10-31