GPU will no longer be the leader in AI chips. Google’s TPU count reaches 15 million in 28 years: for the first time, it beats NVIDIA

Kuai Technology reported on August 1 that in the AI ​​wave in recent years, the one that has benefited the most is NVIDIA. Major companies must purchase the latter’s GPU graphics cards for training and reasoning. However, in two years’ time, GPU may no longer be the dominant one, and Google’s TPU will catch up. On the AI ​​chip technology route, it can be simply divided into general-purpose computing based on NVIDIA-based GPUs and special-purpose computing based on Google’s ASICs. Google’s self-developed tensor processor TPU has a history of 10 years, but the number of deployments cannot compare with the NVIDIA GPUs. However, in the past two years, no one wants the computing power chip to be completely bundled with the NVIDIA graphics card. Self-developed AI chips have begun to explode, and Google’s TPU deployment has also increased rapidly. The seventh-generation TPUv7 has been deployed, and the eighth-generation TPUv8 was released in April, distinguishing training and inference for the first time. Among them, V8T focuses on AI training. Although Google says it can also do inference, it is mainly used for training. Each Pod node is stacked with 9,600 V8T chips, and the FP4 performance reaches 121EFlops. It is equipped with 2PB HBM memory, a memory bandwidth of 19.2TB/s, and an internal chip bandwidth of 400GB/s, which is almost a 2-4 times change. V8i is mainly aimed at AI inference loads, and the specifications have to be reduced a lot. Each node has only 1152 V8i chips, the computing power is reduced to 11.6EFlops, and the memory bandwidth remains unchanged at 19.2TB/s. The next generation of ninth-generation TPUv9 will be even more powerful, with 4-core packaging and higher packaging requirements. Therefore, in addition to the previous cooperation with TSMC, Intel's EMIB packaging will also be used, giving Intel an order for approximately 3 million TPU chips. This also means that the total number of TPUv9 generations is staggering. Recently, supply chain news said that the number of TPUv9 deployed by Google in 28 years will reach 12 million to 15 million. What is this concept? NVIDIA’s AI GPU graphics chips are supplied to most computing power manufacturers on the market. This year’s sales are expected to be 8.2 million units, and will increase to 12.4 million units in 2028.