The question usually arrives in one piece: can I bridge two RTX 5090s and get 64GB for a big model? There are two answers and they point the same way.
Quick answer: no, and you do not need it. NVIDIA’s own specification tables list NVLink (SLI-Ready): No for both the RTX 4090 and the RTX 5090 — the RTX 3090 was the last GeForce card to carry it. More importantly, a bridge was never what let two cards run one large model. Tensor parallelism does that, it works over PCIe, and it is what every serving stack uses.
Check NVIDIA GeForce RTX 5090 on AmazonBuy on Shopee SGWho this is for
You are planning a two-card machine for models that will not fit on one GPU, and you have hit the NVLink question while reading build threads. This page is about what the interconnect does and does not do. For the rest of the build — board, lanes, power — see the multi-GPU setup guide and the dual-GPU PSU guide.
What NVLink is, stated precisely
NVLink is a direct link between two GPUs that is faster than routing the same traffic through the PCIe bus and the CPU. That is the entire feature. It makes communication between cards quicker.
It does not merge their memory. Two 24GB cards joined by a bridge are two 24GB cards that can talk quickly, not one 48GB card. No driver, bridge or setting has ever presented them to an application as a single pool.
This matters because the phrasing people search with — “48GB combined”, “64GB pooled” — describes something that has never existed on consumer hardware. What does exist is better, and it does not need the bridge.
How two cards actually run one large model
Serving frameworks split the model itself across the GPUs. There are two common shapes, and vLLM’s documentation names both:
Tensor parallelism cuts each layer across the cards, so every GPU holds a slice of every layer and they combine results as they go. This is the usual setting for a single machine with two or four cards — in vLLM it is tensor_parallel_size.
Pipeline parallelism assigns whole layers to each card, so the first GPU handles the early layers and hands activations to the second. vLLM recommends it when the GPU count does not divide the model evenly, and it is what llama.cpp does by default when you split a GGUF across devices.
Either way the model’s weights end up distributed, and the usable capacity really is the sum of both cards. A 70B model at Q4_K_M is a 43GB download, which does not fit a single 24GB or 32GB card and does fit across two of them. That has been true on plain PCIe the whole time.
So what does the bridge change?
Communication cost, and only for the shapes that communicate a lot.
Tensor parallelism synchronises at every layer, so it is the sensitive case — a faster link helps most there, and the effect grows with card count and with batch size. Pipeline parallelism passes activations once per boundary, which is far less traffic, so it is largely indifferent to the link.
For a single user typing at a chat window on two cards, the generation speed is set by memory bandwidth and by the model, not by the interconnect. The interconnect starts to show up when you are serving many concurrent requests, or scaling past two cards. If that is your workload, the LLM inference server guide is the better page.
And on current hardware the question is moot in any case: the feature is not on the cards. The practical lever is PCIe — give each card a proper slot rather than a chipset-fed x1, which is covered in the motherboard guide.
Which GPU should you buy for two-card work?
- Want to run 70B at Q4 locally? Two 24GB cards. Two used RTX 3090s remain the cheapest route, and the two-3090 guide covers the build.
- Want one card to do it alone? You are shopping above the consumer line — see how to run 70B on a single GPU for what that actually takes.
- Serving several users at once? Card count and bandwidth both matter, and this is the one case where the interconnect would genuinely help if it existed on your cards.
- Already own a 3090 and want a second? The bridge is a cheap addition and will do no harm. Expect a modest effect on tensor-parallel work and none on anything else.
- Buying new today? Spend the bridge money on PCIe lanes and power headroom instead. There is nothing to attach it to.
Common mistakes
- Buying a bridge for cards that cannot use one. NVIDIA lists NVLink as absent on the 4090 and 5090. The bridges on sale fit the 30 Series.
- Believing memory is pooled. It never was. Frameworks shard the model; the capacity adds up because the weights are split, not because the memory merged.
- Mixing card models and expecting a clean split. Tensor parallelism generally wants matched cards; an uneven pair tends to be limited by the smaller one. Pipeline splitting tolerates mismatch better.
- Putting the second card in an x1 chipset slot. This is the interconnect decision that actually bites on a desktop board, and it costs far more than the absence of a bridge.
- Planning around aggregate VRAM only. Both cards still need room for their share of the KV cache as context grows, and that share is not free. Our VRAM calculator splits weights from cache so you can see it.
Bottom line
| Question | Answer |
|---|---|
| Do current GeForce cards support NVLink? | No — NVIDIA lists it as absent on the 4090 and 5090 |
| Does NVLink combine VRAM into one pool? | No, on any card, ever |
| Can two cards run a model too big for one? | Yes, over ordinary PCIe |
| Does the bridge speed up local LLM inference? | Marginally, for tensor-parallel serving; not for single-user chat |
| What should you optimise instead? | PCIe slot width, power delivery, matched cards |
Two of these still work
NVIDIA GeForce RTX 3090
24GB GDDR6X
24GB each on the used market, and the last GeForce generation that could take a bridge at all — though the pair runs 70B over plain PCIe regardless.
Affiliate links. We may earn a commission at no extra cost to you. Amazon ships worldwide; Shopee SG covers Singapore and ASEAN.
The bridge was never the thing that made two cards hold one model. The software was, and it still is.
Frequently asked questions
Can I combine two GPUs to get more VRAM for a large model?
In the sense that matters, yes: serving frameworks split the model across both cards, so a model too large for either one can run across the pair. In the sense people usually mean, no — the two memories never become a single pool, and no bridge has ever made them one.
Does the RTX 5090 support NVLink?
No. NVIDIA’s published specification table for the card lists NVLink as absent, as it does for the RTX 4090. The last GeForce generation to offer it was the 30 Series.
Is NVLink worth it on two RTX 3090s?
It is a modest, workload-dependent gain rather than a transformation. Tensor-parallel serving with concurrent requests benefits most; a single person chatting with a model will struggle to notice. If you already own the cards it is an inexpensive addition, and nothing breaks without it.
Will PCIe be a bottleneck for two-card inference?
Rarely for generation itself, which is bound by memory bandwidth on each card. It shows up in model loading time and in heavy tensor-parallel serving. The configuration worth avoiding is a second card in a narrow chipset-fed slot.
Do the cards have to be identical?
For tensor parallelism it is strongly preferable, since the work is split evenly and a mismatched pair is held back by the weaker card. Splitting by layers is more forgiving, and people do run mixed pairs that way.
References
- NVIDIA — GeForce RTX 5090 — specification table listing NVLink (SLI-Ready) as absent.
- NVIDIA — GeForce RTX 4090 — the same row, same answer, one generation earlier.
- NVIDIA — GeForce RTX 3090 and 3090 Ti — the last GeForce generation with an NVLink bridge.
- vLLM — parallelism and scaling — tensor parallelism and pipeline parallelism, and when each applies.
- llama.cpp — the layer-splitting path used for GGUF models across devices.
- NVIDIA — NCCL user guide — the collective communication library underneath multi-GPU inference.