Percepta / Accessories
SKU:900-23945-0000-000
NVIDIA's 2-way NVLink bridge for the H200 NVL joins two adjacent cards with 900GB/s of NVLink bandwidth per GPU. One bridge per pair of cards, for servers qualified for the H200 NVL with the NVLink bridge.
Explore the detailsIn stock
Couldn't load pickup availability
Data centre GPU accessory / NVIDIA NVLink
The NVIDIA 2-way NVLink bridge joins two adjacent H200 NVL PCIe cards so they exchange data over NVLink instead of PCIe. NVIDIA rates the link at up to 900GB/s per GPU, against 128GB/s for PCIe Gen5 x16 (both bidirectional). One bridge serves one pair of cards.
This is NVIDIA part 900-23945-0000-000. It is made for the H200 NVL; H100 NVL, H100 PCIe and A100 PCIe cards use a different bridge.
The H100 NVL needs three separate bridges per pair. The H200 NVL instead uses one wide bridge across the top edges of two cards installed side by side. In an eight-GPU server, four 2-way bridges create four NVLink pairs. For a group of four GPUs on NVLink, NVIDIA makes a separate 4-way bridge.
When one model or job is split across both GPUs, for example tensor-parallel LLM inference or multi-GPU training, the cards exchange data frequently. The bridge carries that traffic over NVLink, which NVIDIA rates at up to 900GB/s per GPU (bidirectional). Each card keeps its own 141GB of HBM3e; software that splits the work across both cards can use the faster link between them.
Server vendors qualify their own versions of this 2-way bridge for particular systems: Lenovo 4X67A97320 in the ThinkSystem SR675 V3 and SR650 V4/SR650a V4, HPE S4A90C in the ProLiant DL380a Gen12 and DL385 Gen11, and Cisco UCSX-NVL2-H200 in the UCS X580p PCIe Node. Cisco's version is the NVIDIA part in this listing, 900-23945-0000-000. NVIDIA points to its MGX H200 NVL partner and NVIDIA-Certified Systems. The two cards must sit in adjacent slots.
Specifications for one bridge
| Product | NVIDIA NVLink bridge, 2-way, passive |
|---|---|
| NVIDIA part number | 900-23945-0000-000 |
| Compatible GPU | NVIDIA H200 NVL (PCIe, 141GB HBM3e) |
| Cards linked | Two, installed side by side in adjacent slots |
| Bridges per pair | One |
| NVLink bandwidth | Up to 900GB/s per GPU, bidirectional (NVIDIA rating) |
| Weight | 49g |
| Position | Across the top edges of two adjacent H200 NVL cards |
|---|---|
| Clearance | At least 2.5mm above the cards' top edge, and at least 2.67mm behind the rear card's PCB |
| Recommended topology | Both cards attached to the same CPU or PCIe switch |
| Fitting | Clips onto the cards' NVLink connectors after their covers are removed |
| Removal | Two NVIDIA bridge removal tools, releasing both ends together |
NVIDIA gives up to 900GB/s for NVLink and 128GB/s for PCIe Gen5 x16; both are bidirectional totals per GPU. Server vendors also sell the 2-way bridge under their own part numbers: Lenovo 4X67A97320 (NVIDIA 900-23945-2700-000) and HPE S4A90C. Cisco's UCSX-NVL2-H200 is NVIDIA 900-23945-0000-000, the part in this listing.
Included
One NVIDIA 2-way NVLink bridge for H200 NVL, part number 900-23945-0000-000. Bridge removal tools are not part of this listing.
Plan your installation
One for each pair of H200 NVL cards: one bridge for two cards, two for four cards set up as two pairs, and four for eight cards. To join four cards in a single NVLink group, use NVIDIA's separate 4-way bridge instead.
No. This bridge is for the H200 NVL. The H100 NVL, H100 PCIe and A100 PCIe use a different NVLink bridge, NVIDIA part 900-53651-0000-000, fitted three per pair.
Lenovo, HPE and Cisco each list a 2-way H200 NVL bridge for particular servers: Lenovo 4X67A97320 for the ThinkSystem SR675 V3 and SR650 V4/SR650a V4, HPE S4A90C for the ProLiant DL380a Gen12 and DL385 Gen11, and Cisco UCSX-NVL2-H200 for the UCS X580p PCIe Node. Cisco's part is this NVIDIA part, 900-23945-0000-000. Lenovo and HPE qualify their own part numbers, so check with them before fitting this NVIDIA-branded bridge. The two cards must be in adjacent slots, with at least 2.5mm of clearance above their top edge and at least 2.67mm behind the rear card's PCB.
No. Each H200 NVL keeps its own 141GB of memory and appears as a separate GPU. Software that splits a model or job across both cards, such as tensor-parallel inference or multi-GPU training, can move data between them over NVLink instead of PCIe.
Yes. The bridge can be ordered on its own from this page. For two or more H200 NVL cards, use the enquiry link on the H200 NVL page so we can confirm the order with you.
Discuss the details
Tell us about your application or the system you are putting together.
Accessories
Condition: New
