RTX 5080 + RTX 3090 hits 80 tokens/s on Qwen 3.6 27B Q8 via llama.cpp

User installed an Asus Prime X570-Pro to split PCIe 4.0 lanes into 2x8, disabled CSM, enabled Above 4G Decoding and ReBAR, then loaded NVIDIA driver 610.43.02. llama.cpp was built with GGML_CUDA and multi-GPU flags; nvidia-smi confirmed both cards visible with P2P disabled due to model mismatch. The 5080 handled primary compute while the 3090 supplied additional VRAM, pushing sustained throughput from 50-60 tokens/s (single-card MTP) to 80+ tokens/s.

Benchmark data from the post shows 23646 MiB used on the 3090 and 15861 MiB on the 5080 at idle, with inference scaling directly from added memory bandwidth rather than tensor parallelism. Prior single-card limits on 27B Q8 models stemmed from 16 GB VRAM caps; the 24 GB secondary card removed the quantization ceiling without requiring homogeneous GPUs.

Operationally this configuration lowers the barrier for sustained 70+ token/s local inference on consumer hardware. Different-generation cards still require the closed NVIDIA driver; open-gpu-kernel-modules P2P support remains unavailable. The setup demonstrates that 2026-era 27B-class models now fit comfortably inside two-slot consumer rigs once BIOS and driver constraints are met.

Next measurable threshold is 100 tokens/s on the same model class once llama.cpp adds heterogeneous tensor-parallel scheduling or FP8 kernels land in CUDA 13.x.