NVLink vs GPU Power Connector

NVLink vs GPU Power Connector | Generated by AI

Home 2026.07

Question: How does NVLink physically connect on RTX 3090? Is it related to the GPU power connector (12-pin on RTX 4070)?

Answer:

No. NVLink has nothing to do with the GPU power connector.

The confusion comes from both being “small connectors on the top of the GPU”, but they are completely different.

RTX 4070 power connector

RTX 4070 uses:

12VHPWR (12+4 pin) power connector on many models

It looks like:

PSU cable
    |
    v
[ 12VHPWR ]
+----------------+
| Power input    |
+----------------+

RTX 4070 PCB

Purpose:

PSU ---> GPU

It provides electrical power.

RTX 3090 power connectors

Most RTX 3090 cards use:

8-pin PCIe
8-pin PCIe

or sometimes:

8-pin
8-pin
8-pin

Example:

PSU
 |
 +----8 pin----+
              |
 +----8 pin----+ ---> RTX 3090

RTX 3090 power consumption:

~350W typical
higher under power limit

So it needs more power connectors than RTX 4070.

Where is NVLink on RTX 3090?

RTX 3090 has a separate NVLink finger connector on the top edge of the card.

Example:

                 Top of GPU

        NVLink connector
              ||
              ||
  +-------------------------+
  |                         |
  |       RTX 3090          |
  |                         |
  +-------------------------+

        |             |
     8-pin          8-pin

The NVLink bridge sits here:

RTX 3090 #1

  [ NVLink port ]
          ||
          ||
  [ NVLink bridge ]
          ||
          ||
  [ NVLink port ]

RTX 3090 #2

Important: NVLink is not PCIe power

There are three separate connections:

1. Power

PSU
 |
8-pin connectors
 |
GPU

2. PCIe interface

GPU
 |
PCIe slot
 |
Motherboard

3. NVLink

GPU
 |
NVLink bridge
 |
GPU

They serve different purposes.

RTX 3090 example AI workstation

Two RTX 3090:

             CPU
              |
        PCIe motherboard
        /              \
       /                \
 RTX3090 #1          RTX3090 #2
    |                   |
 8pin+8pin           8pin+8pin
    |
    +==== NVLink ====+

You need:

motherboard with two x16 slots (physical)
enough PCIe lanes
large PSU (850W-1200W)
enough airflow

One more AI-specific note

For LLM training, NVLink bandwidth matters:

RTX 3090:

PCIe 4.0 x16: ~32 GB/s bidirectional theoretical
NVLink: ~112 GB/s bidirectional

So tensor parallel communication can improve.

But modern frameworks:

PyTorch DistributedDataParallel
DeepSpeed ZeRO
FSDP

often work well over PCIe too.

For your situation (local LLM experiments), the main attraction of RTX 3090 is still:

24GB VRAM × 2 = 48GB total capacity

NVLink is a nice optimization, not the main reason to buy it.

Back Donate