Jin Daily AI Trivia: Google’s 7th Gen TPU – Ironwood

Jin Daily AI Trivia: Google’s 7th Gen TPU – Ironwood

Google just announced its next-gen AI chip — and it’s a monster.

⚙️ Raw Power: Each Ironwood chip cranks out roughly 4,614 TFLOPs of FP8 compute. Scale that up to a full superpod with 9,216 interconnected chips, and you’re looking at a staggering 42.5 exaflops of total compute power. (There’s also a smaller 256-chip rack version for developer)

💾 Memory: Each chip packs 192 GB of High Bandwidth Memory (HBM3E) with around 7.4 TB/s of bandwidth

🔋 Efficiency: Ironwood is Google’s most power-efficient TPU yet, delivering 2× the performance per watt compared to its H200-class predecessor, Trillium (TPU v6e).

For context — the only other TPU you can actually buy off the shelf (Cytron Technologies) is the Google Edge TPU from 2018. That tiny 2-watt chip gives you just 4 TOPS of compute power vs 4614 TOPS of ironwood O__O

How times change. 😄

PS: Comparison for fun

1 rack full of ironwood - 64 chip - 294 PFLOPS FP8 - 12.3TB VRAM

1 rack full of NVL72 - 72 chip - 720 PFLOPS FP8 - 13.5TB VRAM