Rack-Scale Architecture with Massive 72-GPU NVLink Domain
-
130 TB/s NVLink aggregate bandwidth across 72 Blackwell GPUs
-
Acts as a single massive GPU - the largest NVLink domain ever offered
- 13.4TB total HBM3E memory (186GB per GPU × 72)
-
Eliminates traditional multi-node communication bottlenecks for trillion-parameter models
30X Real-Time LLM Inference Performance vs H100
-
1.44 exaFLOPS FP4 sparse performance per rack (720 PFLOPS dense)
- Second-generation Transformer Engine with FP4 microscaling
-
Optimized for real-time inference on trillion-parameter models (50ms token-to-token latency)
-
4X faster training at massive scale vs H100 clusters
Integrated CPU+GPU Architecture with Extreme Energy Efficiency
-
36 Grace CPUs + 72 Blackwell GPUs in rack-scale liquid-cooled design
-
25X better performance-per-watt vs air-cooled H100 infrastructure
- 2,592 Arm Neoverse V2 CPU cores with 17TB LPDDR5X memory
-
NVLink-C2C interconnect provides 900GB/s CPU-to-GPU bandwidth per superchip
Get Started
