Mixing NVIDIA and AMD GPUs | Generated by AI

Home 2026.07

Question: Can we mix different NVIDIA GPUs (RTX 4070 + RTX 3090), and can we mix NVIDIA + AMD GPUs in one AI workstation?

Answer:

1. RTX 4070 + RTX 3090: Yes, totally fine

This is a very common setup for AI hobbyists/researchers.

Example:

Machine

CPU
 |
 +-- RTX 3090 24GB  (cuda:0)
 |
 +-- RTX 4070 12GB  (cuda:1)

PyTorch sees:

import torch

print(torch.cuda.device_count())

# 2

for i in range(2):
    print(torch.cuda.get_device_name(i))

Output:

0 NVIDIA GeForce RTX 3090
1 NVIDIA GeForce RTX 4070

CUDA does not require identical GPUs.

But there are trade-offs

Training with DDP

Example:

batch
 |
 +----------+
 |          |
3090       4070

24GB       12GB

The problem:

The smaller GPU limits the batch size.

If your model needs:

GPU0: 20GB
GPU1: 20GB

then:

3090: OK
4070: OOM

The solution:

Use:

uneven batch allocation
model parallelism
ZeRO
FSDP

or simply:

3090 -> training
4070 -> inference / experiments

This is often the best practical setup.

For LLM work, RTX3090 + RTX4070 is actually better than 2×4070

Because:

RTX3090:

24GB VRAM
384-bit memory bus
936 GB/s bandwidth

RTX4070:

12GB VRAM
192-bit memory bus
504 GB/s bandwidth

For LLM:

VRAM matters a lot.

Example:

A 13B model:

FP16:
13B × 2 bytes ≈ 26GB

3090:

24GB
almost fits with quantization

4070:

12GB
needs more compression

2. NVIDIA + AMD GPU together?

Example:

RTX3090
+
AMD MI50 / Radeon Pro VII

Physically:

Yes.

Software:

Much harder.

CUDA world

Most AI frameworks:

PyTorch
 |
CUDA
 |
NVIDIA GPU

NVIDIA path:

PyTorch
 |
CUDA
 |
RTX3090

works perfectly.

AMD path:

PyTorch
 |
ROCm
 |
AMD GPU

is a separate ecosystem.

You cannot do:

one PyTorch process

cuda:0 RTX3090
+
rocm:0 AMD MI50

normally.

They are different backends.

You can run both separately

Example:

Machine:

RTX3090
 |
CUDA
 |
vLLM server


AMD MI50
 |
ROCm
 |
PyTorch experiment

Linux can handle this.

You just run different processes.

Why NVIDIA + AMD is uncommon

The problem is not hardware.

The problem is software stack.

NVIDIA:

CUDA
cuDNN
TensorRT
FlashAttention
xFormers
vLLM
DeepSpeed

AMD:

ROCm
MIOpen
HIP

Many cutting-edge AI libraries optimize NVIDIA first.

Exception: if you like low-level experimentation

For someone like you who likes:

training GPT-2 from scratch
nanoGPT
GPU infrastructure
ROCm exploration

AMD cards are interesting.

Example:

RTX3090:
main LLM work


AMD MI50:
learn ROCm
kernel experiments
cheap compute

This is actually a fun lab.

My ranking for an AI workstation

Best practical

RTX3090 24GB
+
RTX4070 12GB

⭐⭐⭐⭐⭐

Also good

2 × RTX3090

⭐⭐⭐⭐⭐

(but power/heat)

OK

2 × RTX4070

⭐⭐⭐⭐

(good compute, limited VRAM)

Experimental

RTX3090
+
AMD MI50

⭐⭐⭐

(great learning, less convenient)

Avoid

RTX4070
+
AMD consumer GPU

unless you specifically want ROCm.

For a small “AI factory” at home, I would build around one NVIDIA ecosystem first. Mixing NVIDIA + AMD is more like running a mini research lab.

Back Donate