Google Unveils TPU 8 Chips with Enhanced AI Capabilities

Google announced its eighth-generation Tensor Processing Units (TPUs) on April 22, introducing two distinct chip designs: TPU 8t for large-scale model training and TPU 8i for low-latency inference. Both chips are fabricated using TSMC's N3 process with advanced HBM3E memory and will be available to Google Cloud customers later this year. Notably, this marks a shift in Google's strategy as they end Broadcom’s exclusive role in TPU development by bringing MediaTek on board as a silicon design partner.

This dual-chip architecture strategizes to enhance Google's competitive position in AI by catering to specific workload requirements, while also offering alternatives to Nvidia and AMD products. The TPU 8 uses an innovative Boardfly interconnect designed for efficient data transfer, aligning with Google's objective to support high-demand training workloads that favor throughput over raw performance numbers. This move may bolster national AI infrastructure by diversifying chip options in cloud environments, shifting dependency away from Nvidia's technologies.