Google Launches Eighth-Gen TPUs for AI Training

Google has introduced its eighth-generation Tensor Processing Units (TPUs), specifically designed to optimize AI training and inference workloads. The TPU 8t focuses on training, capable of scaling to 9,600 chips with an impressive 121 exaflops of FP4 compute performance and a bandwidth of 19.2Tbps. Its newly architected Virgo Network enhances data transfer capabilities, facilitating the connection of over a million TPU chips and doubling bandwidth for substantial data operations. Meanwhile, TPU 8i caters to inference tasks, offering tailored specifications for low-latency processing.