At Hot Chips 2025, Google launched its next-gen TPU platform Ironwood, boasting remarkable performance breakthroughs that further intensify the AI computing power race.
From TPU v4 to Ironwood, Google achieved a 10x+ leap in single-chip computing power. This 7th-gen TPU delivers 24x the computing capacity of today’s most powerful supercomputers:
- A single Ironwood Superpod integrates 9,216 chips.
- Per-chip peak computing power hits 4,614 TFLOP—over 16x higher than TPU v4 (275 TFLOP) and nearly 10x higher than TPU v5p (459 TFLOP).
- It also packs 192GB HBM memory (7.4TB/s bandwidth) to ensure efficient data processing.
Compared to predecessors (TPU v4 Pod: 4,096 chips; TPU v5p Pod: 8,960 chips), Ironwood breaks new ground in both single-chip performance and system scalability. It uses a modular, expandable architecture:
- Centered on the Ironwood SoC (4 chips integrated on a PCBA motherboard).
- 16 motherboards stack into a rack with 64 chips, then connected via proprietary ICI tech to form a Superpod.
- A high-efficiency liquid cooling system addresses energy consumption and heat dissipation under high computing loads.
Ironwood TPU is Google’s largest and most powerful AI computing engine to date, supporting both AI training and inference—especially optimized for Mixture of Experts (MoE) inference models. Key upgrades include:
- First TPU to support FP8 computing in tensor cores and matrix math units (previous gens only supported INT8 inference and BF16 training).
- Equipped with 3rd-gen SparseCore accelerators (to speed up embedding operations for recommendation models).
ICgoodFind: We’ll keep tracking cutting-edge trends in AI chips as Google Ironwood fuels the AI computing power sector.
