✌️NVLink

About NVLink

NVLink is supported only on instance flavors that include 8× NVIDIA H100 or H200 GPUs. In this configuration, NVLink provides high-bandwidth, low-latency GPU-to-GPU communication, enabling:

  • Faster model training, especially for large models that require frequent inter-GPU data exchange

  • Improved scaling efficiency when using distributed training frameworks (e.g., Megatron-LM, DeepSpeed, PyTorch FSDP)

  • Reduced communication bottlenecks compared to PCIe-only GPU connectivity

  • Higher overall compute throughput for workloads that depend on multi-GPU synchronization

Instances with fewer than 8 GPUs do not support NVLink.

Enable NVLink for 8x GPUs Flavors

Note: NVLink is not enabled by default in FPT-provided images. Users must manually enable NVLink if required.

To enable NVLink support, follow the steps below:

  • Open the file: /etc/default/grub.d/00-fci-grub.cfg

  • Locate the following line and remove or comment it out:

GRUB_CMDLINE_LINUX_DEFAULT="nvidia.NVreg_NvLinkDisable=1"

  • Update the GRUB configuration:

sudo update-grub

Here is a clean, well-structured user-guide rewrite:

Verification

To verify that NVLink has been enabled successfully, run the following commands.

1. Check NVLink status

The output should show active NVLink connections with their link speeds (e.g., 25 GB/s per lane).

2. Check GPU topology

The output should display NVLink (NV##) connections between all GPUs, similar to the example below:

Legend

  • X : Same GPU

  • SYS : PCIe + inter-NUMA interconnect

  • NODE: PCIe + interconnect within one NUMA node

  • PHB : PCIe Host Bridge

  • PXB : Multiple PCIe bridges

  • PIX : Single PCIe bridge

  • NV# : NVLink connection with # bonded links

Troubleshooting

If you encounter errors such as:

  • system not yet initialized

  • Issues when calling torch.cuda.device_count() or torch.cuda.get_device_name(i)

Restart the NVIDIA Fabric Manager:

Then retry your application or verification steps.

PreviousBlock StorageNextFAQ

Last updated