Mixing a 9070 XT with an RTX 3080 for local LLMs: 30-50 tok/s and hard lessons
BrightCandle · reddit · 2026-10-19
A Redditor documents running Qwen3.8-27B on a 9070 XT + RTX 3080 combo. Key lessons: never plug monitors into the NVIDIA card (Windows makes it primary and performance craters), layer split beats tensor split due to the 3080's PCIE 4.0 x4 bottleneck, and MTP in llama.cpp roughly doubles throughput. Final result: 30-50 tok/s on Qwen3.8 28B Q4K-M with GPU utilization at only 60%/30%.
More from Infra
- Anthropic Built Scion, Its Own Source Control System, as Merged Changes Grew 12x to 30,000/Day — Pavan_Belagatti · 2026-10-19
- Dwarfstar on 128GB M5 Max Mac Studio hits 16 tok/s with SSD offload via DeepSeek Flash 4.1 — antirez · 2026-10-19
- HBM shake-up: SK Hynix HBM4 packaging delays reported, prices seen up 30-45% in 2027 — tengyanAI · 2026-10-19
- Researcher warns of ~1,000 cloud inference endpoints potentially compromised in active risk — matthew_d_green · 2026-10-19
- AWS open-sources TorchNeuron, a native PyTorch backend that makes Trainium a standard PyTorch device — PyTorch · 2026-10-19
- Container machines: persistent Linux dev VMs unique to Apple's container runtime — AIFlow_ML · 2026-10-19