NVLink with Two RTX 2080 Ti

NVLink with Two RTX 2080 Ti | Generated by AI

Home 2026.08

Question: Can we use NVLink with two RTX 2080 Ti 22GB cards?

Answer:

Yes. Two RTX 2080 Ti cards can use NVLink, including modified 22GB VRAM versions, because NVLink support depends on the GPU model and PCB, not the VRAM size.

NVIDIA GeForce RTX 2080 Ti supports 2-way NVLink.

Hardware setup

You need:

RTX 2080 Ti #1
       ||
       ||  NVLink bridge
       ||
RTX 2080 Ti #2

The NVLink bridge connects the two GPUs directly.

Specs:

NVLink generation: Turing NVLink
Bandwidth: ~50 GB/s bidirectional
Maximum GPUs: 2 (consumer RTX 2080 Ti)
PCIe slots still required

But: NVLink does NOT automatically give 44GB VRAM

This is the important part.

Many people think:

22GB + 22GB = 44GB VRAM

No. For normal CUDA applications:

GPU0:
  22GB VRAM

GPU1:
  22GB VRAM

They remain separate memory pools.

Example:

torch.cuda.device_count()

# 2

torch.cuda.get_device_properties(0).total_memory
# 22GB

torch.cuda.get_device_properties(1).total_memory
# 22GB

PyTorch sees:

cuda:0 -> 22GB
cuda:1 -> 22GB

For LLM inference

Frameworks can split the model:

Example:

Llama 70B

Layer 0-35
   GPU0 (22GB)

Layer 36-70
   GPU1 (22GB)

Tools:

vLLM tensor parallel
HuggingFace Accelerate
DeepSpeed inference
llama.cpp multi-GPU

Example:

python -m vllm.entrypoints.openai.api_server \
    --model Qwen/Qwen3-30B \
    --tensor-parallel-size 2

The model is sharded across GPUs.

For training

NVLink helps because GPUs communicate faster:

Without NVLink:

GPU0
 |
 PCIe
 |
GPU1

With NVLink:

GPU0
 ||
 || 50GB/s
 ||
GPU1

Useful for:

DDP
tensor parallel
pipeline parallel

But RTX 2080 Ti has no:

NVLink memory pooling
unified VRAM
NVSwitch

Compare with RTX 3090

Two RTX 2080 Ti 22GB:

VRAM:
22 + 22 = 44GB

Compute:
2 × 13.4 TFLOPS FP32
≈ 26.8 TFLOPS

Two RTX 3090:

VRAM:
24 + 24 = 48GB

Compute:
2 × 35.6 TFLOPS FP32
≈ 71 TFLOPS

3090 also has:

Ampere Tensor Cores
BF16 support
much faster transformer kernels

For your nanoGPT / vLLM experiments

A pair of 2080 Ti 22GB is actually interesting:

2 × RTX 2080 Ti 22GB

VRAM:
44GB

Good for:
✓ QLoRA 7B/14B
✓ inference of 30B-40B quantized models
✓ distributed training experiments
✓ learning tensor parallel

But:

2 × RTX 3090 24GB

is the sweet spot for a home AI lab because:

48GB VRAM
much faster kernels
BF16
FlashAttention compatibility

The 2080 Ti 22GB is mainly attractive if you find them extremely cheap. For learning LLM systems engineering, two GPUs are valuable because you can experiment with the same ideas used in large clusters: tensor parallelism, pipeline parallelism, NCCL communication, and distributed training.

Back Donate