Redesigning the Inference Chip: From Nvidia GPU's Flaws to OpenAI Jalapeño