When Is NVLink Worth It? A Real-World Test with RTX 3090s
I tested NVLink on dual RTX 3090s to see if the $300 bridge is worth it for AI workloads. While my server setup with full PCIe lanes made NVLink unnecessary, I found it crucial for consumer platforms where bandwidth drops to x8 or x4. FSDP training saw massive gains with NVLink, whereas standard inference remained largely unaffected unless you skip the unofficial P2P driver patch.
These products are hilariously named and marketed, but alas, consumer CPU limitations mean they cannot take full advantage of multiple GPUs, which is often the main way to upgrade an AI workflow.
- colingauvin
Great piece. Was wondering whether to get an NVLink bridge for $500+ for my two A4500s for inference. This pretty much confirms it only really rescues you if you have nonsensical PCI-E topology.
- deflator
Very informative article. I was hazy on where NVLink sat in the local inference stack. This makes it very clear: to adapt consumer hardware that lacks multiple full x16 PCIe lanes.
- causality0
I badly miss the days when I could buy a GPU and then three years later buy the same GPU for a hundred bucks and boost my framerate by 70%.