NVIDIA KILLER? Google Built a 9,216-Chip Liquid-Cooled 'AI Hypercomputer скачать в хорошем качестве

NVIDIA KILLER? Google Built a 9,216-Chip Liquid-Cooled 'AI Hypercomputer

An extensive analysis of Google's coordinated hardware strategy to vertically integrate its cloud and AI infrastructure, referred to as a "two-pronged assault" on the "Age of Inference." This strategy centers on the introduction of two custom chips: the Ironwood Tensor Processing Unit (TPU), an ASIC designed to challenge Nvidia's dominance by achieving per-chip specification parity and optimizing for continuous AI inference workloads, and the Axion CPU, an Arm-based processor intended to replace third-party x86 hardware for general-purpose tasks to improve Google Cloud's profit margins. The document emphasizes that the true product is the "AI Hypercomputer," a massive, 9,216-chip system that uses a proprietary Optical Circuit Switching (OCS) fabric to enable petabyte-scale memory sharing and high reliability. Finally, the analysis covers the sophisticated software stack (JAX/Pallas and PyTorch integration) meant to neutralize the CUDA ecosystem, details the competitive landscape against Nvidia, and highlights the crucial, multi-billion-dollar role of Broadcom (AVGO) as Google's indispensable custom-silicon manufacturing partner.