35B Model Performance Plummets on Mixed GPUs
alchninja · reddit · 2026-07-15
The author attempted to run Qwen3.5 35B-A3B in a mixed GPU + CPU environment on a machine with a Ryzen 5700X, 32GB RAM, an RTX 5070 Ti 16GB, and a newly added RTX 3060 12GB.
Previously, running 35B Q6KXL on a single GPU (with -ncmoe 30, V quantized to Q8) yielded around 750pp / 55tg, which was fast enough for daily use. However, after adding the second card to accommodate longer context and MTP, switching to -ncmoe22 left both cards with only about 512MB of free VRAM, causing performance to crash to 130pp / 5tg. Despite trying different split modes and settings that don't degrade the model's "intelligence," they couldn't reach the desired 600pp / 30-35tg, and asked if there was an issue with the configuration.
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