Buy another RTX 3090 for NVLink

Buy another RTX 3090 for NVLink | Generated by AI

Home 2026.07

Question: If using NVLink, what GPU should you buy instead of adding an RTX 3060?

Answer:

For your i9-13900KF + RTX 4070 system, NVLink changes the answer completely.

First: RTX 4070 does not support NVLink.

The RTX 40 series consumer cards removed NVLink:

GPU	NVLink support
RTX 3090	✅ Yes
RTX 3090 Ti	❌ No
RTX 4090	❌ No
RTX 4080	❌ No
RTX 4070	❌ No
RTX 3060	❌ No

So you cannot do:

RTX 4070 + RTX 3060 + NVLink

There is no connector.

If you want NVLink, buy another RTX 3090

The realistic second-hand AI setup is:

RTX 3090 24GB
+
RTX 3090 24GB

with NVLink bridge.

You get:

GPU 0:
24GB GDDR6X

GPU 1:
24GB GDDR6X

Total:
48GB VRAM

NVLink provides a high-bandwidth GPU-to-GPU connection:

PCIe:
CPU
 |
PCIe x16
 |
GPU

GPU <-- PCIe --> GPU

versus:

NVLink:

GPU0 <=======> GPU1
      high bandwidth

This helps for:

model parallel inference
some distributed training
tensor parallel workloads

But there is an important catch

NVLink does not magically merge VRAM.

This:

2 x RTX 3090

24GB + 24GB

does not become:

one CUDA device with 48GB

CUDA sees:

torch.cuda.device_count()

2

You need software support:

DeepSpeed ZeRO
Megatron-LM tensor parallel
vLLM tensor parallel
HuggingFace Accelerate

Example:

vllm serve Qwen2.5-32B \
 --tensor-parallel-size 2

then:

GPU0:
layers 0-?

GPU1:
layers ?-?

Compared with RTX 4070 + RTX 3060

Your current:

RTX 4070
12GB
+
RTX 3060
12GB

= 24GB total

Cheap, but:

no NVLink
mixed GPU generations
slower communication

3090 pair:

RTX 3090
24GB
+
RTX 3090
24GB

= 48GB

Much better for LLM work.

Power requirement

Your current PSU:

750W

is not enough.

3090:

RTX 3090:
350W

two:
700W

CPU:
250W

system:
100W

Peak:

~1000W

You want:

1200W PSU

minimum.

Examples:

Corsair HX1200
Seasonic PRIME TX-1300
Super Flower Leadex 1200W

Motherboard requirement

For dual RTX 3090:

You need:

PCIe slot spacing:

GPU1
|
3-slot gap
|
GPU2

Most normal ATX boards are too cramped.

Better:

ASUS ProArt Z790-Creator WiFi
Gigabyte Z790 AORUS Master
MSI MEG Z790 ACE

These have wider spacing.

For your specific situation

Given you already have RTX 4070:

I would not buy RTX 3060.

The choices:

Option A — cheapest AI expansion

RTX 4070
+
RTX 3090

Pros:

36GB total VRAM
CUDA
one modern + one large VRAM card

Cons:

no NVLink
mixed GPUs

This is probably the sweet spot.

Option B — serious local LLM workstation

Sell RTX 4070:

2 x RTX 3090 24GB
NVLink

You get:

48GB VRAM

Can run:

Llama 70B quantized
Qwen 72B quantized
larger context inference

Option C — keep RTX 4070

Honestly, for your workload (nanoGPT training, LoRA, agent development), I would do:

Current:
RTX 4070 12GB

Add:
RTX 3090 24GB

Total:

36GB VRAM

Use:

RTX 3090 → LLM inference/training
RTX 4070 → experiments/display

This gives much more value than RTX 3060.

The RTX 3090 second-hand market exists exactly because it is still one of the cheapest ways to get 24GB CUDA VRAM.

Back Donate