Google announces TPU 8t and TPU 8i at Cloud Next, splitting training and inference

Google DeepMind/Project Suncatcher

Google announced the eighth-generation TPUs as two specialized chips: TPU 8t (Sunfish) optimized for training with Broadcom, and TPU 8i (Zebrafish) for inference with MediaTek. This marks a shift from the unified design of past generations. Expected availability is late 2027 using TSMC's 2nm process, delivering nearly 3 times the on-chip memory per chip for speed.

• TPU 8t (Sunfish) specializes in model training workloads.
• TPU 8i (Zebrafish) specializes in serving inference queries.
• Design reflects that training and inference have different performance requirements.
• Built on TSMC's 2nm process, expected in late 2027.
• Delivers 3x on-chip memory versus previous generation.