How to Select and Buy Cables for NVIDIA SXM GPU AI Servers and NVIDIA GPU Clusters
When companies build NVIDIA GPU clusters, most attention goes to the servers and GPUs. But the cables and switches are just as important.
Modern 8-GPU AI servers, such as NVIDIA HGX systems with SXM GPUs, can generate huge network traffic. If the cables are wrong, the cluster can lose performance, become unstable, or be delayed during installation. This is especially true for NVIDIA H100, H200, B200, B300 and similar GPU servers using NVIDIA 400G or 800G networking.
This guide explains how to choose the right NVIDIA AI server cables for 8-GPU servers and GPU cluster infrastructure.
NVIDIA GPU Cluster Cables and Transceivers
Pricing and Availability on Request
Why NVIDIA GPU Cluster Cables Matter
In AI training, HPC and large inference workloads, NVIDIA GPU servers constantly exchange data. The network must support high bandwidth and low latency.
The cable infrastructure affects:
GPU utilization
Training speed
Latency
Cluster stability
Rack airflow
Troubleshooting
Future expansion
Total project cost
For NVIDIA GPU clusters, cables are not just accessories. They are part of the performance design.
Typical Setup and Cables for NVIDIA 8-GPU AI Servers
A modern NVIDIA 8-GPU server may use several high-speed network adapters, not just one cable.
For example, the NVIDIA DGX H100 and DGX H200 use eight single-port ConnectX-7 cards for the compute network, at up to 400Gb/s each, plus two dual-port ConnectX-7 cards for storage and in-band management. OEM 8-GPU servers such as the Dell PowerEdge XE9680 with H200 SXM5 GPUs also need several high-speed network links per server.
Common NVIDIA networking infrastructure includes:
400G NDR InfiniBand
400GbE Ethernet
800G links (twin-port 2×400G NDR, or native 800G XDR)
NVIDIA ConnectX-7 or ConnectX-8 adapters
NVIDIA Quantum or Spectrum switches
OSFP, OSFP112, QSFP112 or QSFP-DD ports
Native and breakout (splitter) cables
A cluster with 8, 16 or more GPU servers can quickly require dozens or hundreds of NVIDIA 400G cables or 800G cables. This is why cable planning must happen before installation.
Main Types of NVIDIA AI Server Cables
NVIDIA DAC cables
DAC means Direct Attach Copper.
DAC cables are good for short connections inside the same rack. They are usually cheaper than optical cables, but at 400G or 800G speeds they can be thick, heavy and harder to manage.
For NVIDIA 400G and 800G (100G-PAM4) networks, DAC cables are offered from 0.5 to 2 meters for straight cables and up to 3 meters for splitters, with twin-port OSFP, OSFP and QSFP112 ends. See our guide to 800G OSFP DAC InfiniBand cables.
Use DAC when:
The server and switch are very close
Cable length is short
Cost is important
Cable density is not too high
For large NVIDIA GPU clusters, DAC is often not the best choice for every link.
NVIDIA active copper cables (ACC / LACC)
ACC means Active Copper Cable. NVIDIA calls its current version LACC, which stands for Linear Active Copper Cable.
LACC cables extend copper links from 3 to 5 meters, which can span several racks. They are DAC cables with an additional IC in each end that boosts the signal. NVIDIA offers them in twin-port OSFP for 800G-to-800G switch-to-switch links at 3, 4 and 5 meters. The 1:2 and 1:4 LACC splitters have been discontinued. More details are in our guide to 800G OSFP ACC InfiniBand cables.
Use ACC / LACC when:
The link is too long for DAC but still within a few meters
You want the low power and low latency of copper
You connect switch to switch with twin-port OSFP
AOC cables in NVIDIA GPU clusters
AOC means Active Optical Cable.
AOC cables were very common in EDR and HDR (100G and 200G) NVIDIA InfiniBand clusters because they are lighter and easier to route than copper cables.
For NDR 400G and 800G networks, this has changed. NVIDIA does not offer AOCs for its 100G-PAM4 (400G/800G) InfiniBand or Ethernet products, because the OSFP ends are heavy and very likely to break the fibers during installation. The only exception is a 1:2 AOC splitter, up to 30 meters, for backward compatibility between NDR or 400GbE switches and HDR or 200GbE equipment.
Use AOC when:
You connect NDR or 400GbE switches to HDR or 200GbE switches and adapters
You expand an older EDR or HDR cluster
For new NDR and XDR clusters, use DAC or LACC for short links, and transceivers with fiber for longer links.
NVIDIA optical transceivers with fiber cables
Another option is using separate NVIDIA optical transceivers and fiber patch cables.
This is useful for:
Longer distances
Structured cabling
Patch panels
Connections between rows or rooms
Data center designs that need more flexibility
For NVIDIA NDR and XDR clusters, this is the main option for links longer than a few meters. NVIDIA's 800G twin-port OSFP transceivers include 2×SR4 multimode (up to 50 meters), 2×DR4 single-mode (up to 100 or 500 meters) and 2×FR4 single-mode (up to 2 km).
This option gives more flexibility, but it is also more complex. You must match the transceiver, fiber type, connector, speed, switch and NIC.
The table below compares the main cable types for NVIDIA 400G and 800G networks:
Cable type | Reach (NVIDIA 400G/800G) | Typical use |
DAC | 0.5–2 m straight, up to 3 m splitters | Short links inside one rack |
LACC (ACC) | 3–5 m | Switch-to-switch links across nearby racks |
AOC | Up to 30 m, backward-compatible 1:2 splitters only | NDR or 400GbE to HDR or 200GbE equipment |
Transceivers with fiber | Up to 50 m (2×SR4), 100 or 500 m (2×DR4), 2 km (2×FR4) | Longer links, rows, rooms and structured cabling |
Native NVIDIA 800G Cables
A native NVIDIA 800G cable connects one 800G port directly to another 800G port.
Be careful with the term "800G". On NVIDIA Quantum-2 NDR switches such as the QM9700, each OSFP cage contains two NDR ports of 400G. An 800G twin-port OSFP cable between two NDR switches carries 2×400G, not one native 800G link. Native 800G per port comes with XDR InfiniBand, on NVIDIA Quantum-X800 switches and ConnectX-8 SuperNICs. Read more in our comparison of NDR 400G vs XDR 800G InfiniBand.
Example:
800G twin-port OSFP to twin-port OSFP DAC or LACC (2×400G NDR, switch to switch)
Use straight NVIDIA 800G cables when both sides use the same 800G port type. See our guide to 800G OSFP to OSFP cables.
Before buying, check:
Connector type
Switch port type
NIC port type
Finned-top or flat-top OSFP
Speed support
InfiniBand or Ethernet mode
Cable length
Firmware and compatibility
Do not assume that every NVIDIA 800G cable works with every 800G port.
NVIDIA 800G to 2×400G Breakout Cables
NVIDIA breakout cables are also called splitter cables.
A common example in NVIDIA GPU clusters is:
800G twin-port OSFP to 2× 400G OSFP splitter
This allows one 800G twin-port switch cage to connect to two 400G server-side ports, if the switch and NIC support this mode. On the QM9700, each NDR port can also be split into two lower-speed ports, for example 200Gb/s NDR200.
With NVIDIA 400G and 800G parts, short breakout links can use DAC splitters up to 3 meters. The 1:2 and 1:4 QSFP112 DAC splitters and the LACC splitters have been discontinued, so longer breakout links use transceivers with 1:2 fiber splitter cables.
Breakout cables help with:
Higher port density
Better use of switch ports
Connecting many 400G server ports
Large 8-GPU server clusters
Cleaner infrastructure design
But breakout cables must be checked carefully. The switch, NIC, firmware and cable must all support the same breakout mode.
A native NVIDIA 800G cable and an 800G to 2×400G breakout cable are not the same product. Our guide to 800G OSFP to 2×400G breakout cables explains the options.
Connectors for NVIDIA Cables: OSFP, OSFP112, QSFP112 and QSFP-DD
Connector type is one of the easiest mistakes to make when buying NVIDIA cables.
High-speed NVIDIA GPU infrastructure may use:
OSFP
OSFP112
QSFP112
QSFP-DD
QSFP56
QSFP28
The cable must match both sides.
Examples:
Twin-port OSFP to twin-port OSFP (switch to switch)
Twin-port OSFP to 2× OSFP breakout (switch to ConnectX-7 OSFP adapters)
Twin-port OSFP to 2× QSFP112 with transceivers and fiber (switch to ConnectX-7 QSFP112 adapters or BlueField-3)
QSFP112 to QSFP112
Two OSFP details cause many wrong orders:
Twin-port or single-port: NVIDIA switches use the twin-port 800G (2×400G) OSFP. ConnectX-7 OSFP adapters use a single-port 400G OSFP. Single-port 400G OSFP or QSFP112 devices cannot be used in twin-port OSFP switch cages.
Finned-top or flat-top: Air-cooled Quantum-2 and Spectrum-4 SN5600 switches use finned-top (IHS) twin-port OSFP. ConnectX-7 OSFP adapters only support riding heat sink (RHS) cages, so they need flat-top OSFP ends. Read our guide to OSFP finned-top vs flat-top cables.
Some suppliers use the name OSFP112 for OSFP parts with 100G-PAM4 lanes. NVIDIA documentation uses OSFP and twin-port OSFP, so always compare the exact part number, not only the name.
Do not buy only by speed. An "NVIDIA 400G cable" or "NVIDIA 800G cable" is not enough information.
NVIDIA InfiniBand Cables or Ethernet Cables?
Many NVIDIA GPU training clusters use InfiniBand because of low latency and strong RDMA performance. Our guide to InfiniBand networking for NVIDIA H100 and H200 GPU clusters explains why.
Other clusters use Ethernet, especially when the customer wants standard data center networking or AI Ethernet designs.
Before buying NVIDIA InfiniBand cables or Ethernet cables, confirm:
Is the network InfiniBand or Ethernet?
Which switch is used?
Which NIC is used?
Is the adapter configured for InfiniBand or Ethernet?
Does the cable support the required mode?
Some adapters can support both InfiniBand and Ethernet, but the full path must be correct: NIC, cable, switch, firmware and software.
NVIDIA Switches and Cables: Check More Than the Model
GPU clusters often use high-speed switches such as NVIDIA Quantum-2 QM9700 for NDR InfiniBand or similar platforms. The QM9700 has 32 OSFP cages with 64 NDR 400Gb/s ports. See our overview of NVIDIA QM9700, QM9790 and Quantum-X800 InfiniBand switches.
When buying NVIDIA switches and cables together, check:
Exact switch model
Port speed
Port count
Connector type
Breakout support
Firmware
Power supplies
Rail kit
Airflow direction
The exact switch version matters, especially in dense NVIDIA GPU racks.
Airflow Direction Matters for NVIDIA Switches and Cables
Airflow is not a small detail.
NVIDIA switches can be:
Front-to-rear airflow
Rear-to-front airflow
For example, the QM9700 is specified for both forward and reverse airflow, and each version has its own part number.
NVIDIA AI servers may have high-speed network ports on the front or rear. If the airflow direction does not match the rack design, installation becomes harder and cooling can suffer.
Always confirm airflow before ordering switches.
Also check cable routing. Too many thick or badly routed cables can block airflow and make service work harder.
Length Planning for NVIDIA GPU Cluster Cables
Do not guess cable lengths.
Common NVIDIA GPU cluster cable lengths include:
1m
2m
3m
5m
7m
10m
12m
15m
20m
With NVIDIA 400G and 800G parts, copper covers only the shortest of these lengths: up to 2 meters for straight DAC, up to 3 meters for DAC splitters and 3 to 5 meters for LACC. Longer links use transceivers with fiber.
The right length depends on:
Server position
Switch position
Same rack or multi-rack layout
Front or rear cabling
Cable management
Bend radius
Service access
Too short can create physical stress. Too long creates mess, blocks airflow and increases cost.
For larger projects, prepare a cable list by length and quantity before ordering.
Can Existing NVIDIA Cables Be Reused?
In NVIDIA GPU cluster expansion projects, some existing cables may be reused.
This can reduce cost, but the counting must be very careful.
Check:
Cable type
Length
Connector
Speed
Native or breakout
Condition
Compatibility with the new layout
Do not assume all old NVIDIA cables can be reused. Sometimes only part of the previous batch fits the new design.
Buying Refurbished, Used or Excess NVIDIA GPU Cluster Cables
New NVIDIA 400G and 800G cables can be expensive and hard to source quickly. Refurbished, used or excess stock NVIDIA cables can be a good option if they are tested and correctly identified.
Before buying, check:
Exact part number
Manufacturer
Cable type
Connector type
Finned-top or flat-top OSFP
Speed
Length
Native or breakout
InfiniBand or Ethernet support
Switch compatibility
NIC compatibility
Quantity available
Same-batch availability
Test status
Warranty
Return conditions
For NVIDIA GPU clusters, buying one cable is easy. Buying 64, 128 or 256 correct cables is much harder.
Common Mistakes When Buying NVIDIA GPU Cluster Cables
Avoid these mistakes:
Buying only by speed
Mixing up native and breakout cables
Treating a twin-port 800G NDR cable as a native 800G XDR link
Ignoring connector type
Mixing up finned-top and flat-top OSFP
Planning AOCs for NDR or XDR links
Forgetting InfiniBand vs Ethernet
Not checking switch airflow
Guessing cable lengths
Mixing too many cable models
Not buying spare cables
Ordering before the rack layout is clear
Not checking what existing cables can be reused
These mistakes can delay the full NVIDIA GPU cluster installation.
What to Send When Requesting a Quote for NVIDIA Cables
To get the right NVIDIA cable offer, send:
NVIDIA GPU server details
Server model
GPU type
Number of servers
Number of GPUs per server
NIC model
Number of NIC ports
NIC connector type
NVIDIA switch details
Switch model
Number of switches
Port speed
Connector type
Airflow direction
InfiniBand or Ethernet
NVIDIA cable details
Native or breakout
Speed
Connector type
Finned-top or flat-top OSFP
Length
Quantity per length
New, refurbished or excess stock
Required lead time
Warranty expectation
Rack layout for the NVIDIA cables
Same rack or multi-rack
Switch position
Server position
Front or rear cabling
Spare cable requirement
With this information, a supplier can check the correct options much faster. You can send it to us through our request a quote form.
FAQ: NVIDIA GPU Cluster Cables for 8-GPU AI Servers
What cables are used for NVIDIA 8-GPU AI servers?
NVIDIA 8-GPU AI servers usually use high-speed DAC, active copper (LACC), breakout cables, or optical transceivers with fiber. With NVIDIA NDR and XDR networks, AOCs are only used for backward compatibility with HDR or 200GbE equipment.
Is one NVIDIA 400G cable enough for an 8-GPU server?
Usually no, because many NVIDIA 8-GPU servers use several 400G or 800G network links.
What is a native cable in NVIDIA GPU clusters?
A native cable connects one high-speed port directly to another port, for example 800G to 800G.
What is a breakout cable in NVIDIA GPU clusters?
A breakout cable splits one high-speed port into multiple lower-speed links, for example 800G to 2×400G.
Is an NVIDIA 800G to 2×400G breakout cable the same as an 800G cable?
No, they are different products and must be matched to the switch, NIC, and firmware.
Should I use DAC or AOC cables for NVIDIA GPU clusters?
For NVIDIA 400G and 800G networks, the choice is usually DAC, LACC or transceivers with fiber, not AOC. DAC is good for short same-rack links up to 2 or 3 meters, LACC covers 3 to 5 meters, and transceivers with fiber are used for longer distances. NVIDIA does not offer AOCs for its 400G/800G (100G-PAM4) products, except backward-compatible splitters to HDR or 200GbE.
Are NVIDIA OSFP and QSFP112 cables the same?
No, OSFP and QSFP112 are different connector types and are not physically interchangeable.
What is OSFP112, and how does NVIDIA name it?
OSFP112 is a name some suppliers use for OSFP modules and cables with 100G-PAM4 lanes, used for 400G and 800G AI networking. NVIDIA documentation uses OSFP and twin-port OSFP instead.
What is the difference between finned-top and flat-top NVIDIA OSFP cables?
Finned-top (IHS) OSFP ends have cooling fins and are used in air-cooled NVIDIA switches. Flat-top (RHS) ends have no fins and are used in ConnectX-7 OSFP adapters and systems with riding heat sinks, such as DGX H100 and H200.
Can I use the same NVIDIA cable for InfiniBand and Ethernet?
Sometimes yes, but only if the cable, NIC, switch, firmware, and configuration support it.
What cables are used with NVIDIA Quantum-2 QM9700 switches?
QM9700 switches use finned-top twin-port OSFP cables and transceivers: DAC and LACC for short links, transceivers with fiber for longer links, and native or breakout (splitter) options. NVIDIA does not offer AOCs for NDR, except backward-compatible splitters to HDR or 200GbE.
Why does NVIDIA switch airflow direction matter?
Airflow direction matters because it affects cooling, cable routing, and rack installation quality.
How do I calculate NVIDIA GPU cluster cable lengths?
Calculate cable lengths from the real rack layout, including server position, switch position, cable path, and service access.
Can I reuse existing NVIDIA GPU cluster cables?
Yes, but only if the old cables match the new speed, connector, length, native/breakout type, and layout.
Should I buy spare cables for an NVIDIA GPU cluster?
Yes, spare cables are recommended because one wrong or failed cable can delay the full NVIDIA GPU cluster installation.
Is it safe to buy refurbished NVIDIA GPU cluster cables?
Yes, if they are tested, correctly identified, compatible, and sold with warranty.
Why are NVIDIA 400G and 800G cables expensive?
They are expensive because NVIDIA 400G and 800G cables are high-speed components built for strict performance and compatibility.
What information should I send before requesting a quote for NVIDIA cables?
Send the server model, GPU type, NIC, switch, protocol, connector, cable type, lengths, quantities, airflow, and rack layout.
What is the most common mistake when buying NVIDIA GPU cluster cables?
The most common mistake is buying by speed only, without checking connector type, protocol, breakout design, and compatibility.
Final Thoughts on NVIDIA GPU Cluster Cables
For NVIDIA 8-GPU AI servers, cable selection is part of the cluster design.
The right cables help the GPUs communicate at high speed, reduce installation problems, improve airflow and make the cluster easier to support.
The wrong cables can delay the project, create compatibility issues and waste expensive GPU capacity.
Before buying NVIDIA GPU cluster cables, check the full infrastructure: servers, GPUs, NICs, switches, protocol, connectors, cable lengths, airflow and rack layout.
If you are building or expanding an NVIDIA AI cluster, we can help you check the correct NVIDIA cables, switches, quantities, lengths and available refurbished or excess stock options.
Send us your switch model, NIC model, cable lengths and quantities through our request a quote form, and we will help you prepare a correct cable plan for your NVIDIA GPU cluster.
Sources: NVIDIA Cables and Networking Documentation
NVIDIA GPU Cluster Cables and Transceivers
Pricing and Availability on Request






Comments