5070 Ti, RX 9070 build hits 100 tps

Cross-vendor VRAM pooling via Vulkan is now a viable, high-performance alternative to the "CUDA tax" for local LLM inference. The RX 9070’s 256-bit bus and 645 GB/s bandwidth offer superior performance-per-dollar compared to a second NVIDIA card, while Vulkan's maturation in llama.cpp allows seamless memory pooling across NVIDIA and AMD hardware without significant overhead. Strategic task splitting and the use of models like Qwen 3.6 35B represent a new sweet spot for high-speed 32GB setups.