NVIDIA NVLink Bridge Overview

NVIDIA NVLink Bridge Overview | Generated by AI

Home 2026.07

Question: What is NVIDIA NVLink Bridge (2-slot / 3-slot / 2-way / 4-way) used for, especially for RTX 3090, A6000, A100, H100?

Answer:

NVLink Bridge is a physical connector that links two NVIDIA GPUs together with a high-speed GPU-to-GPU interconnect. It is not an adapter that combines GPUs into one giant GPU automatically. The software (CUDA, PyTorch, NCCL, etc.) must support multi-GPU communication. (NVIDIA)

Think of it like:

Without NVLink:

GPU 0 ---- PCIe ---- CPU ---- PCIe ---- GPU 1

With NVLink:

GPU 0 ===== NVLink ===== GPU 1

The second path has much higher GPU-to-GPU bandwidth and lower latency.

1. What does “2-slot / 3-slot” mean?

It describes the physical distance between the two GPUs on the motherboard.

Example:

2-slot bridge

GPU
[====]

(space)

GPU
[====]

The GPUs are separated by 2 PCIe slots.

3-slot bridge

GPU
[====]

(space)

(empty)

GPU
[====]

The GPUs are farther apart.

The bridge is just a rigid connector. You need the correct spacing.

2. What does “2-way / 4-way” mean?

2-way NVLink

Most consumer/workstation cards:

GPU0 <==== NVLink Bridge ====> GPU1

Examples:

RTX 3090
RTX A6000
RTX 6000 Ada

They support pairing two GPUs. (NVIDIA)

4-way NVLink

This is mainly server/HPC territory:

GPU0 ===== GPU1
 ||         ||
GPU2 ===== GPU3

Usually requires:

SXM modules
NVSwitch
HGX/DGX systems

Example:

A100 SXM
H100 SXM

A normal desktop motherboard with PCIe RTX cards cannot build a 4-way NVLink fabric. (NVIDIA Developer)

3. RTX 3090 NVLink Bridge

RTX 3090 is special because it is the last GeForce card with NVLink.

Two RTX 3090:

RTX3090 24GB
      |
   NVLink
      |
RTX3090 24GB

Memory:

48GB total VRAM

but not like a single 48GB GPU.

For example:

Good:

distributed training
some CUDA workloads
model parallelism

Not automatic:

model.to("cuda")

does NOT become 48GB.

PyTorch still sees:

cuda:0 -> 24GB
cuda:1 -> 24GB

You need:

DDP
FSDP
DeepSpeed
tensor parallel frameworks

4. RTX 3090 + A6000 bridge compatibility

Interesting point:

The RTX A6000 NVLink bridges are often used for RTX 3090 because Ampere NVLink uses the same generation. Community testing confirmed A6000 2-slot/3-slot bridges can work with RTX 3090. (Reddit)

Example:

RTX3090
   |
A6000 3-slot NVLink Bridge
   |
RTX3090

This is common in AI workstation builds.

5. A100 / H100 are different

A100 PCIe

Similar idea:

A100 PCIe
    |
 NVLink
    |
A100 PCIe

Two-card setups exist.

A100 SXM / H100 SXM

These are datacenter accelerators:

        NVSwitch
       /   |   \
H100  H100 H100 H100

They have much stronger interconnect.

H100 uses 4th generation NVLink:

H100: 900 GB/s total NVLink bandwidth
A100: 600 GB/s total NVLink bandwidth (NVIDIA Developer)

6. For your AI workstation use case

Given your interest in local LLM training/inference:

Dual RTX 3090

Good:

2 x RTX3090
=
48GB VRAM
+
NVLink

Useful for:

30B~70B quantized models
LoRA training
distributed inference

But:

NVLink does not magically make:

single GPU 48GB

Many modern LLM frameworks use PCIe/NCCL efficiently anyway.

Buying priority

For AI:

VRAM
  >
GPU compute
  >
memory bandwidth
  >
NVLink

Usually:

2 x RTX3090 without NVLink

can still beat

1 x RTX3090 with NVLink

because you get:

48GB physical VRAM
more CUDA cores
more throughput

NVLink is a bonus.

Simple summary

Product	NVLink?	Typical
RTX 3060	❌	no
RTX 4070	❌	no
RTX 3090	✅	2 GPU bridge
RTX A6000	✅	2 GPU bridge
A100 PCIe	✅	2 GPU
A100 SXM	✅✅	NVSwitch systems
H100 SXM	✅✅	NVSwitch systems

For a home AI lab, the interesting one is:

2 × RTX3090 24GB
+
A6000 2-slot/3-slot NVLink bridge

but only after you have enough PCIe lanes, PSU capacity, and cooling. (Reddit)

References:

Back Donate