What Is NVL36?_
The NVL36 platform is a dense GPU compute node designed for high-performance AI workloads, integrating 36 GPUs in a single rack enclosure. It leverages NVLink for high-bandwidth, low-latency GPU-to-GPU communication, enabling efficient scaling of model parallelism. This configuration is typically used in data centers requiring large-scale training or inference clusters.
Technical Details
NVL36 uses copper-based NVLink spine connections inside the rack to link all 36 GPUs, avoiding fiber for intra-rack GPU communication. The platform supports liquid cooling to manage thermal loads, with coolant distribution units (CDUs) supplying the necessary flow. Power delivery is handled via high-current busbars, and networking for scale-out (e.g., InfiniBand or Ethernet) uses separate MPO fiber connections. Follow OEM guidelines for rack mounting; coolant connectors typically use push-pull mechanisms without torque requirements.
How Leviathan Systems Works with NVL36
In our field work, NVL36 racks arrive as pre-assembled units that we integrate into the data center floor, connecting structured cabling for both liquid cooling loops and fiber networking. We commission the NVLink copper spine by verifying link integrity per OEM specs, and we ensure the CDU is properly paired with the rack's thermal load.
Related service
GPU Rack Assembly & Rack-and-Stack →