Testing llama.cpp PCIe Bandwidth on Dual GPUs
nick_ziv · reddit · 2026-07-12
This post shares PCIe transfer measurements for dual-GPU local inference using llama.cpp, aiming to determine if multi-GPU setups can use a 1x riser without significantly degrading performance.
Test Environment
- Ubuntu 24.04, single-socket Xeon motherboard
- Two GPUs: RTX 3090 24GB (250W) and Titan RTX 24GB (225W)
- Model: Qwen3.6-27B-UD-Q4KXL.gguf, context length 180k
- PCIe bandwidth monitored via nvidia-smi dmon -s t
Results Comparison
Tensor parallel:
- Continuous inter-GPU transfers during decode/inference
- Higher decode speed, around 47 t/s
- Lower prefill speed, around 650 t/s
Pipeline parallel (llama.cpp default):
- Faster prefill, around 1250 t/s
- Lower transfer during decode/inference
- Decode speed around 30 t/s
Conclusion
The author concludes:
- Tensor parallel is better for boosting decode
- Pipeline parallel's prefill is closer to single-GPU capability
- PCIe 3.0 x16 doesn't appear fully saturated in this test, making a 1x riser potentially viable
More from Infra
- China’s AI arms race is increasingly defined by chips, data centers, and open models — BenBajarin · 2026-07-22
- Agent search bottlenecks are now about variance, not raw latency — rohanpaul_ai · 2026-07-22
- Gavin Baker argues Nvidia may be one of open source AI’s biggest supporters — GavinSBaker · 2026-07-22
- AI Power Demand Exposes US Energy Gap, Urging Shift from Scarcity to Abundance — bradneuberg · 2026-07-22
- Gavin Baker says Nvidia’s $630B figure would be system revenue, not all Nvidia’s — GavinSBaker · 2026-07-22
- A Firecracker-based platform says it can host 6,000 AI agents on one 256 GB server — maritime_sh · 2026-07-22