|
Quantity
|
Out of stock
|
||
|
|
|||
|
|
|||
Why bridge two cards at all
Two cards in one server talk over the PCIe bus by default, and every exchange between them passes through the CPU. An NVLink bridge lays a direct line between the cards, with bandwidth several times higher than PCIe and noticeably lower latency.
The practical consequence is the one that matters: card memory stops being two separate pools. Two 141 GB cards under a bridge give 282 GB, enough to hold the model whole instead of splitting it and shuttling pieces back and forth.
When a bridge genuinely helps
- The model does not fit one card. This is the main case. Without a bridge you have to shard it, and the exchanges eat the speed.
- Training and fine-tuning. Gradients are reduced across cards constantly, and here the bandwidth gap turns into hours.
- One job across several cards. If each card handles its own independent request, a bridge adds almost nothing.
What to check before buying
A bridge is matched to the specific card model and to the slot spacing: two-slot and four-slot versions are not interchangeable. The cards have to sit next to each other and be identical. The motherboard has to lay out the slots so the NVLink connectors line up.
Send us the card and motherboard models and we will tell you which bridge fits and whether it will pay off in your particular workload.