NVIDIA Blackwell
NVL72 rack-scale coherent GPU domains, the default substrate for frontier training and inference.
Full NVIDIA Blackwell record →Head to head
NVIDIA Blackwell vs Google TPU Ironwood
NVIDIA Blackwell (NVIDIA) and Google TPU Ironwood (Google) both sit in Chips & Compute. Google TPU Ironwood shipped more tracked changes in the last 30 days (12 vs 9). Every value below comes from the latest crawl of the vendors' own pages.
| Attribute | NVIDIA Blackwell NVIDIA | Google TPU Ironwood |
|---|---|---|
| Segment | Chips & Compute | Chips & Compute |
| Version | Blackwell | Ironwood (7th generation) |
| Entry cost | Pricing is not publicly listed (enterprise hardware architecture/system) | from $5.4/unit |
| Pricing tiers | — | On-Demand (us-central1 Iowa) $12 · DWS Flex-start price (us-central1 Iowa) $6 · DWS Calendar Mode price (us-central1 Iowa) $8.4 · 1-year Commitment (us-central1 Iowa) $8.4 · 3-year Commitment (us-central1 Iowa) $5.4 · On-Demand (europe-west2 London) $13.2 · DWS Flex-start price (europe-west2 London) $6 · DWS Calendar Mode price (europe-west2 London) $8.4 |
| Model stack | Blackwell · Blackwell B200 · Grace CPU | Gemini · Ironwood · Ironwood (TPU v7) · Ironwood TPU · Ironwood TPU (7th Generation) · TPU v7 · TPU7x · TPU7x (Ironwood) · XLA compiler |
| Context window | — | — |
| Public API | ||
| Routes models | ||
| Changes / 30d | 9 | 12 |
| Origin | Global | Global |
| Capabilities | 72-GPU NVLink domain, FP4 inference, HBM3E 192GB, CUDA ecosystem, Second-generation Transformer Engine, Fifth-generation NVLink, FP4 AI Support, Tensor Cores with microscaling formats | 9,216-chip pods, inference optimised, liquid cooled, JAX + PyTorch/XLA, Large-scale training, Reasoning and inference, 9,216 liquid-cooled chips per pod, 42.5 ExaFlops performance |
Google TPU Ironwood
v7 TPU pods scaling to 9,216 chips, available through Google Cloud and used for Gemini.
Full Google TPU Ironwood record →