Do I need NVLink for distributed training?

No — you do not need NVLink for distributed training.

Short version

• Distributed training works over standard PCIe, Ethernet, InfiniBand, or other interconnects.
• NVLink is optional and mainly helps with faster GPU-to-GPU communication on the same node.
• You can absolutely do multi-GPU / multi-node training without NVLink.

When NVLink helps

NVLink can improve performance if your workload has a lot of:

• gradient synchronization
• tensor/model parallel communication
• activation transfers between GPUs
• GPU memory pooling / peer-to-peer access

This is especially useful for:

• large models
• model parallelism
• highly communication-bound training

When you don’t need it

If you’re doing:

• data parallel training
• small-to-medium models
• multi-node training over normal networking
• single-GPU training

then NVLink is not required.

What matters more than NVLink

For distributed training, the key factors are usually:

• GPU count and memory
• interconnect bandwidth/latency
• network quality between nodes
• training framework support (PyTorch DDP, FSDP, DeepSpeed, Horovod, etc.)

Practical takeaway

If you have NVLink, great — it can improve speed.
If you don’t, you can still train distributed models normally; you may just see lower performance in communication-heavy setups.

If you want, I can also tell you when NVLink actually makes a measurable difference for your specific training setup.