NVIDIA ConnectX-7 vs ConnectX-8: What Optical Transceivers, DAC and AOC Do AI Networks Need?

As AI servers become more powerful, the network interface connecting those servers to the fabric must evolve as well.

NVIDIA ConnectX-7 has become widely used for 200G and 400G AI, HPC and data center connectivity, while the newer ConnectX-8 generation moves the architecture toward PCIe 6.0, 400GbE-class Ethernet and 800Gb/s XDR InfiniBand connectivity.

For optical transceiver buyers, however, the most important question is not simply:

“ConnectX-7 or ConnectX-8?”

The more useful questions are:

What is the NIC port type?

Is it OSFP or QSFP112?

Is the required link 200G, 400G or 800G-class?

Is the application Ethernet or InfiniBand?

How far does the link need to reach?

These questions determine whether the correct solution should be a 200G / 400G optical transceiver, 800G-class OSFP connectivity, DAC, AOC or breakout solution.

For AI server manufacturers, system integrators and data center operators, understanding this relationship is essential before purchasing optics.

ConnectX-7 vs ConnectX-8 at a Glance

The comparison image highlights several commonly used ConnectX-7 and ConnectX-8 configurations.

A simplified view is:

NIC Example	Network Capability	Port Type	PCIe	Optical Connectivity Opportunity
ConnectX-7 MCX75310AAS-NEAT	400GbE / NDR 400Gb/s	1× OSFP	PCIe 5.0 x16	400G OSFP
ConnectX-7 MCX715105AS-WEAT	400GbE / NDR 400Gb/s	1× QSFP112	PCIe 5.0 x16	400G QSFP112
ConnectX-7 MCX755106AS-HEAT	2×200GbE / NDR200	2× QSFP112	PCIe 5.0 x16	200G QSFP-family connectivity
ConnectX-7 MCX755106AC-HEAT	2×200GbE / NDR200	2× QSFP112	PCIe 5.0 x16	200G QSFP-family connectivity
ConnectX-7 MCX753436MC-HEAB	2×200GbE / NDR200	2× QSFP112	PCIe 5.0 x16, OCP 3.0	200G connectivity
ConnectX-8 C8180 / 900-9X81E-00EX-ST0	2×400GbE or 800Gb/s XDR	1× OSFP	PCIe 6.0 x16	800G-class OSFP / 2×400G connectivity
ConnectX-8 C8240 / 900-9X81Q-00CN-ST0	2×400GbE / NDR	2× QSFP112	PCIe 6.0 x16	Dual 400G QSFP112

NVIDIA's current documentation confirms that the ConnectX-7 MCX75310AAS-NEAT supports 400GbE or NDR 400Gb/s through a single OSFP port, while MCX715105AS-WEAT uses a single QSFP112 port for 400GbE/NDR. The dual-port MCX755106AS-HEAT and MCX755106AC-HEAT operate at 200GbE / NDR200 per port.

ConnectX-7: A Key Platform for 200G and 400G AI Networking

ConnectX-7 remains highly relevant for AI servers, GPU clusters, HPC and high-performance Ethernet or InfiniBand systems.

Depending on the exact adapter model, ConnectX-7 can use:

OSFP

or

QSFP112

interfaces.

This is extremely important for optical buyers because two NICs with the same 400G network speed may require different physical optical form factors.

For example:

400GbE + OSFP

and

400GbE + QSFP112

are not the same module.

The speed may be identical, but the module housing and electrical interface are different.

ConnectX-7 with 400G OSFP

One of the important examples is:

MCX75310AAS-NEAT

NVIDIA specifies this model with:

400GbE Ethernet

or

400Gb/s NDR InfiniBand

through:

1 × OSFP

with PCIe Gen 5 support.

For an optical transceiver supplier, this creates demand for:

400G OSFP optical transceivers

as well as:

400G OSFP DAC

and potentially:

400G OSFP AOC

depending on deployment distance.

Which 400G OSFP Optical Module Should You Choose?

The answer depends mainly on distance and fiber infrastructure.

A simplified selection may look like:

Short multimode link
→ 400G OSFP SR4

500m-class single-mode link
→ 400G OSFP DR4

Longer single-mode link
→ FR-class or other appropriate optical solution

For example, if a customer tells Optech:

ConnectX-7 OSFP, 400G, 500m

the likely direction is:

400G OSFP DR4 + SMF

rather than SR4.

But the exact switch on the other side must also be checked.

ConnectX-7 with 400G QSFP112

The image also includes:

MCX715105AS-WEAT

This adapter supports:

400GbE / NDR 400Gb/s

through:

1 × QSFP112

according to NVIDIA's current documentation.

This changes the optical requirement.

Instead of:

400G OSFP

the NIC-side solution becomes:

400G QSFP112

Typical optical options may include:

400G QSFP112 SR4

400G QSFP112 DR4

depending on distance and fiber type.

This is why customers should always provide the exact NIC part number, not simply say:

“We are using ConnectX-7.”

Same ConnectX-7, Different Optical Module

This is one of the most important lessons for optical transceiver buyers.

Two customers may both say:

“We use ConnectX-7 at 400G.”

But one may need:

400G OSFP

while another requires:

400G QSFP112

Therefore:

NIC family alone is not enough.

The correct selection process is:

NIC Model
→ Exact P/N
→ Port Form Factor
→ Speed
→ Protocol
→ Distance
→ Optical Module

ConnectX-7 Dual 200G QSFP112 Configurations

The image also includes ConnectX-7 models such as:

MCX755106AS-HEAT

and

MCX755106AC-HEAT

These use:

2 × QSFP112

and support:

200GbE per port

or NDR200 operation.

NVIDIA's official documentation confirms these dual-port 200G configurations.

For optical connectivity projects, this means one NIC can generate demand for:

two 200G links

instead of one 400G link.

Depending on the target switch and required distance, those links may use:

• 200G optical transceivers

• DAC

• AOC

• breakout connectivity

ConnectX-7 OCP 3.0 for AI Servers

The supplied comparison also shows:

MCX753436MC-HEAB

which uses the OCP 3.0 form factor.

NVIDIA documents this adapter as a dual-port QSFP112 ConnectX-7 OCP 3.0 card supporting 200GbE or NDR200, with PCIe 5.0 x16 and hardware cryptography support.

OCP NICs are particularly relevant in hyperscale and server platforms where NICs are integrated into standardized server architectures.

For Optech, the optical requirement remains driven primarily by:

port type + link speed + reach

rather than whether the NIC itself is PCIe add-in card or OCP.

ConnectX-8: Moving Toward 800G-Class AI Connectivity

ConnectX-8 represents a major step forward for next-generation AI and HPC infrastructure.

Compared with ConnectX-7, one of the most visible changes is:

PCIe 6.0 x16

along with support for more advanced high-speed network configurations.

The ConnectX-8 C8180, model:

900-9X81E-00EX-ST0

uses a:

single OSFP cage

and supports:

800Gb/s XDR InfiniBand

or:

2 × 400GbE

depending on configuration.

This creates an important new generation of optical connectivity requirements.

ConnectX-8 C8180: One OSFP Cage, Multiple Network Modes

This is a concept buyers need to understand carefully.

The C8180 has:

one physical OSFP cage

but NVIDIA supports several logical port configurations, including:

1 × 800Gb/s XDR InfiniBand

2 × 400GbE Ethernet

and additional split configurations.

Therefore, the correct optical or cable solution depends on how the card is configured.

For an XDR application, the buyer may require an 800G-class OSFP interconnect architecture.

For Ethernet, the same card can operate as:

2 × 400GbE

which may require dual-channel or breakout-oriented connectivity.

This is another reason why buying optics only by looking at the metal cage is not enough.

ConnectX-8 C8240: Dual QSFP112 for 2×400G

The second major ConnectX-8 example shown is:

900-9X81Q-00CN-ST0

also known as:

C8240

NVIDIA documents this product as:

Dual-port QSFP112

with:

400GbE / NDR capability per port

and PCIe 6.0 x16.

This creates a very straightforward optical opportunity:

Port 1 → 400G optical connection

Port 2 → 400G optical connection

A fully populated C8240 can therefore require:

2 × 400G QSFP112 connectivity per NIC.

At data center scale, the optical module quantity can increase quickly.

Why ConnectX-8 Matters to Optical Transceiver Suppliers

AI network bandwidth is rising because GPU systems are exchanging larger amounts of data.

When one server moves from:

200G

to:

400G

or:

2×400G

the number and bandwidth of required optical links increases accordingly.

At cluster scale, hundreds or thousands of NICs may translate into thousands of high-speed optical links.

This creates demand for:

400G QSFP112

400G OSFP

800G-class OSFP

DAC

AOC

and:

breakout connectivity

depending on the architecture.

ConnectX-7 vs ConnectX-8: What Changes for Optical Buyers?

The most important changes can be summarized as follows.

ConnectX-7

Strongly associated with:

200G / 400G

PCIe 5.0

OSFP and QSFP112 options

NDR / NDR200

400GbE and 200GbE

ConnectX-8

Moves toward:

400G per port

and:

800Gb/s XDR-class connectivity

PCIe 6.0

OSFP and QSFP112 configurations

higher-bandwidth AI and HPC fabrics

For optical buyers, this means the transition from ConnectX-7 to ConnectX-8 can create a migration from:

200G → 400G

and increasingly:

400G → 800G-class connectivity

Which Optical Module Should ConnectX-7 Customers Buy?

A practical decision process is:

ConnectX-7 + OSFP + 400G

Consider:

400G OSFP optical transceiver

or:

400G OSFP DAC / AOC

ConnectX-7 + QSFP112 + 400G

Consider:

400G QSFP112 optical transceiver

or:

QSFP112 DAC / AOC

ConnectX-7 + Dual QSFP112 + 200G

Consider:

200G connectivity per port

based on the target switch, distance and host qualification.

Which Optical Module Should ConnectX-8 Customers Buy?

For ConnectX-8:

C8180 Single OSFP

First confirm whether the card is configured as:

800Gb/s XDR

or:

2×400GbE

before choosing the optical solution.

C8240 Dual QSFP112

The optical requirement is typically centered on:

400G QSFP112 connectivity per port.

Again, the link partner must also be checked.

SR, DR or FR: Distance Still Matters

After confirming NIC and port form factor, the next question is:

How far is the connection?

A typical decision pattern is:

Short Reach

Use SR-class optics where appropriate.

Often suitable for multimode fiber inside racks or data halls.

500m-Class

DR-class optics are frequently used for high-speed single-mode data center connections.

Longer Reach

FR-class or other longer-reach optical architectures may be more suitable.

Therefore, the final product decision is not:

ConnectX-8 = one specific optical module.

It is:

ConnectX-8 + port type + speed + distance + fiber + switch = correct optical module.

NIC-to-Leaf Connectivity in AI Clusters

One of the most common applications is:

GPU Server
↓
ConnectX NIC
↓
Optical Transceiver / DAC / AOC
↓
Leaf Switch

For short rack-level connections:

DAC

can provide low power and cost.

For longer integrated links:

AOC

may be suitable.

For structured fiber infrastructure:

Optical Transceiver + Fiber

offers greater flexibility.

Straight vs Breakout Connectivity

High-speed AI systems also increasingly use breakout links.

Examples include:

800G → 2×400G

or:

400G → 2×200G

depending on the host platforms and port configuration.

This makes it especially important to understand the switch side.

For example:

ConnectX-7 400G NIC

may connect to:

an 800G switch port operating as 2×400G

rather than a dedicated 400G physical switch port.

In that case, the switch-side module and NIC-side module may have different form factors but matching optical interfaces.

OSFP Does Not Always Mean the Same Thing

Another common mistake is assuming:

OSFP on one device automatically equals OSFP on the other.

That is not necessarily true.

A switch may use:

800G twin-port OSFP

while a NIC uses:

400G single-port OSFP

The mechanical family is similar, but the lane architecture and operating mode can be different.

Professional optical selection therefore requires checking both host devices.

What Customers Should Send Optech

For fast and accurate product selection, customers should provide:

• NIC model

• Exact NVIDIA P/N

• Switch model

• Required Ethernet or InfiniBand mode

• Required speed

• Port form factor

• Link distance

• Fiber type

• Connector type

• Straight or breakout topology

• Required quantity

For example:

NIC: ConnectX-7 MCX75310AAS-NEAT
Port: OSFP
Speed: 400G
Distance: 500m
Fiber: SMF
Switch: target model provided

Optech can then evaluate the correct 400G OSFP optical solution.

Why Choose Optech for ConnectX-7 and ConnectX-8 Connectivity?

Optech Technology provides high-speed optical and cable solutions for AI, HPC, cloud and data center networks.

Our portfolio includes:

200G optical connectivity

400G OSFP

400G QSFP112

800G OSFP

DAC

AOC

ACC / AEC

Breakout solutions

and next-generation high-speed interconnect products.

For NVIDIA NIC projects, Optech can support customers with:

Product Selection

Matching the correct module to NIC, switch, distance and fiber.

Compatibility Evaluation

Testing products against target host platforms where applicable.

Sample Qualification

Allowing customers to validate optics before larger deployments.

Customized Coding and Labeling

Supporting project-specific requirements.

Volume Supply

For data center, AI server, system integration and channel customers.

From NIC Model to the Correct Optical Module

The recommended optical selection process is:

Host Equipment

→ Exact NIC P/N

→ OSFP / QSFP112

→ 200G / 400G / 800G

→ Ethernet / InfiniBand

→ Distance

→ MMF / SMF

→ SR / DR / FR

→ Straight / Breakout

→ Optech Product

This approach helps prevent incorrect module selection and reduces compatibility risk.

FAQ

What optical module does ConnectX-7 use?

It depends on the exact ConnectX-7 model. Some 400G models use OSFP, while others use QSFP112. Dual-port models may operate at 200GbE per QSFP112 port.

Does every ConnectX-7 use OSFP?

No.

ConnectX-7 is available with different interfaces, including OSFP and QSFP112 depending on the specific adapter.

Can ConnectX-7 support 400GbE?

Yes. NVIDIA documents multiple ConnectX-7 configurations supporting 400GbE, including both OSFP and QSFP112 variants.

Does ConnectX-8 support 800G?

ConnectX-8 C8180 supports an 800Gb/s XDR InfiniBand mode through a single OSFP cage. For Ethernet, NVIDIA documents 2×400GbE configurations rather than describing that configuration simply as one 800GbE Ethernet port.

Does ConnectX-8 support QSFP112?

Yes. The C8240 uses dual QSFP112 ports and supports 400GbE / NDR-class connectivity.

Should I choose OSFP or QSFP112?

Choose according to the exact NIC hardware.

Do not select the module based only on speed.

What is best for a 400G 500m connection?

A DR-class single-mode optical solution is commonly considered for 500m-class applications, but the exact form factor and host compatibility must be confirmed.

Can ConnectX NICs use DAC?

Yes, appropriate DAC solutions can be used for short-distance high-speed connections when supported by the host and link partner.

Can ConnectX NICs use AOC?

AOC can be useful for longer integrated links than passive copper, subject to compatibility and application requirements.

What information does Optech need before recommending a module?

Please provide the exact NIC P/N, switch model, speed, protocol, distance, fiber, connector and quantity.

Conclusion

The transition from ConnectX-7 to ConnectX-8 is not only a NIC upgrade.

It represents a broader shift in AI infrastructure from:

200G and 400G

toward:

400G and 800G-class networking.

ConnectX-7 already creates significant demand for 400G OSFP, 400G QSFP112 and 200G connectivity, while ConnectX-8 introduces PCIe 6.0 and new high-bandwidth configurations such as 800Gb/s XDR over OSFP and dual 400GbE connectivity.

For optical buyers, the correct purchasing logic is:

Do not buy by speed alone.

Instead, confirm:

NIC → Port → Protocol → Speed → Distance → Fiber → Module.

For companies deploying NVIDIA ConnectX-7 or ConnectX-8 in AI, HPC, GPU server or data center networks, Optech can support 200G, 400G and 800G optical transceivers, DAC, AOC, breakout solutions, compatibility evaluation, samples and volume orders.

Contact Optech with your NVIDIA NIC P/N, switch model and link distance for product recommendations and project pricing.