4 min read RackmountNTS offers a liquid-cooled, rack-scale platform with 72 NVIDIA B200 GPUs and 36 Grace CPUs. It is built for AI training, inference, and HPC. These workloads need dense GPU compute in one integrated rack.
Liquid cooling keeps thermals stable under sustained load. That helps data centers run high-power GPU clusters with steady performance. It can also extend hardware life.
Overview of the Rack-Scale Platform
This solution pairs NVIDIA Blackwell GPU compute with Grace CPU capacity. It targets teams that need high parallelism for large models, simulation, and analytics.
Core elements include:
- 72 NVIDIA B200 GPUs for large-scale parallel processing
- 36 Grace CPUs for host-side compute and coordination
- Direct liquid cooling for dense thermal management
- Flexible storage and expansion for enterprise data sets
The design fits AI factories and HPC sites. It scales compute without spreading jobs across many disconnected racks.
Key Hardware: GPUs and Grace CPUs
The 72 B200 GPUs provide the main compute for AI and HPC. They suit training, fine-tuning, and high-throughput inference when models and batches are large.
The 36 Grace CPUs support orchestration and data prep. They also help workloads that need fast host compute next to the GPU complex. Together, GPUs and CPUs form a balanced rack for data-heavy jobs.
Hardware highlights:
- High GPU count for dense compute in one rack
- Grace CPUs aligned with NVIDIA accelerated platforms
- Design aimed at large-model and scientific workflows
Teams building mixed clusters can pair this rack with other NTS GPU server nodes.
Liquid Cooling and Thermal Management
Air alone struggles when many high-TDP GPUs run at full load. Liquid cooling removes heat closer to the source. That supports quieter, more stable operation.
Benefits of liquid cooling in this platform include:
- Better heat removal under sustained AI training loads
- Lower reliance on high-speed fan arrays
- More consistent GPU clocks with less thermal throttling
- Improved component life in dense layouts
Good thermal design matters as rack power climbs. Liquid cooling helps operators use facility power more productively.
Storage, Memory, and Expansion
AI and HPC jobs create large datasets, checkpoints, and logs. This platform supports storage that matches enterprise needs.
Typical capabilities include:
- Scalable DDR5 memory for fast host-side work
- Hot-swap storage for service with less downtime
- Expansion paths for drives or companion storage nodes
Pair storage with NTS vault and enterprise chassis when you need high-bay tiers for long-term data.
Performance for AI, HPC, and Data Center Workloads
This rack-scale system targets jobs that stress compute and I/O:
- Large language model training and fine-tuning
- Deep learning for vision, speech, and multimodal models
- Scientific simulation and engineering analytics
- Real-time inference at scale for enterprise AI services
High GPU density helps finish training cycles faster. It can also serve more concurrent inference requests. Grace CPUs help keep data pipelines and control tasks responsive.
Integration, Scalability, and Long-Term Planning
The platform fits data centers that already support liquid-cooled racks. Modular growth lets you add capacity as models and demand grow.
Planning considerations include:
- Facility readiness for liquid cooling manifolds and CDUs
- Network bandwidth for multi-GPU and multi-rack traffic
- Power delivery and backup for sustained GPU load
- Software stack consistency across training and inference nodes
Blackwell-compatible designs help protect investments as GPU generations advance.
Use Cases and Application Scenarios
This solution fits teams that need rack-scale GPU density. It reduces the need to manage dozens of separate servers by hand.
- AI research labs training foundation and domain models
- Enterprise AI platforms serving internal copilots and analytics
- Defense and federal programs that need secure, high-performance compute
- Media and simulation workloads with heavy parallel math
Each case benefits from integrated cooling, dense GPUs, and enterprise build options from RackmountNTS.
Related NTS Platforms
Rack-scale B200 deployments often sit with other NTS systems:
- NTS Elite Fusion 6U liquid-cooled GPU server for dense AI and HPC nodes
- NTS Elite APEX 8U GPU servers for flexible 8-GPU training and inference
- Enterprise storage chassis for datasets, checkpoints, and archive tiers
Your RackmountNTS team can map compute, storage, and networking into one cluster design.
Conclusion
This liquid-cooled rack-scale platform delivers dense AI and HPC compute. It uses 72 NVIDIA B200 GPUs and 36 Grace CPUs. Liquid cooling and high GPU count address limits of traditional air-cooled racks.
Contact RackmountNTS to discuss facility needs, cluster design, and deployment options.
