Rack-Scale Grace Blackwell Platform
NVIDIA GB200 NVL72 on AltusCloud
GB200 NVL72 is designed for AI factory-scale deployments, integrating Grace CPUs and Blackwell GPUs into a rack-scale architecture optimized for trillion-parameter model inference and large-scale training.
Highlights
Rack-scale acceleration for AI factories
Up to 30x vs H100 systems
Real-time LLM inference
Up to 4x faster
LLM training at scale
130 TB/s rack-scale
NVLink domain bandwidth
Blackwell rack-scale architecture
GB200 NVL72 combines 72 Blackwell GPUs in a massive NVLink domain, enabling high-throughput, low-latency communication for large distributed model workloads. This architecture is designed to reduce bottlenecks in trillion-parameter systems.
Real-time inference at scale
The GB200 profile is aimed at low-latency inference for very large models, where communication efficiency and memory bandwidth are critical for throughput.
Massive-scale training
For organizations running AI factory programs, GB200 offers strong scaling behavior for training and refinement workflows across large infrastructure footprints.
Networking and operations layer
AltusCloud supports GB200 deployments with enterprise networking options, cluster orchestration workflows, and operations guidance for sustained production performance across AI factory environments.
Specifications
| Metric | GB200 NVL72 Class |
|---|---|
| Configuration | 36 Grace CPUs + 72 Blackwell GPUs |
| FP4 Tensor (sparse) | 1,440 PFLOPS |
| FP8/FP6 Tensor (sparse) | 720 PFLOPS |
| FP16/BF16 Tensor (sparse) | 360 PFLOPS |
| FP32 | 5,760 TFLOPS |
| GPU Memory / Bandwidth | 13.4 TB HBM3e / 576 TB/s |
| NVLink Bandwidth | 130 TB/s |
| CPU Core Count | 2,592 Arm Neoverse V2 cores |
| CPU Memory / Bandwidth | 17 TB LPDDR5X / up to 14 TB/s |
Specifications are reference-level and deployment plans are finalized during solution design.
Ready to Deploy
Deploy NVIDIA GB200 with AltusCloud
Contact our infrastructure team to plan cluster sizing, region strategy, and enterprise purchasing for your AI platform.