tl;dr: Author combined an RTX 5080 (16GB) and a refurbished RTX 3090 (24GB) on an Asus Prime X570-Pro to run Qwen 3.6 27B at Q8 with a 230k context across 39GB of VRAM. Key setup details include enabling Above 4G Decoding/ReBAR, disabling CSM, building llama.cpp with `CMAKE_CUDA_ARCHITECTURES="86;120"` and NCCL off, and using tensor split mode with MTP+ngram speculative decoding. The result: 80-90 tokens/sec generation, with PCIe running at x8/x8 Gen4.