Google splits its eighth-generation TPU into training and inference chips

Google says the training chip delivers 2.8x the performance of Ironwood for the same price and inference is 80% better. Both were promised for later in 2026.

Why it matters: Separate chips for training and inference signal how large the inference market has become.

Performance

2.8x Ironwood at the same price (Google)

Inference

80% better (Google)

SRAM

384 MB per chip, triple Ironwood

Availability

Promised later in 2026