​ This is south Korean start up all-in on inference chip: https://furiosa.ai/renegade-spec Tsmc 5nm node Hynix HBM3 1.5TB/s 48GB VRAM TDP 180W Already tested on LG LLM. If they opened their programming interface the way NVIDIA opens PTX and Intel opens SPIR-V, and team up with llama.cpp for getting a GGML backend working, it would be a game changer. Rtx pro 5000 48gb (non-hbm) is $5k now. Amd's r9700 32gb is $1.3k Intel B70 32gb is $1k I bet if their RNGD chip is ...