Is NVLink Worth It for Local LLMs? Dual RTX 3090 (August 2026)

Bottom Line (August 2026)

• Running Ollama, LM Studio or llama.cpp? Skip NVLink. Layer-split inference barely touches the interconnect. Every guide that says “you do not need NVLink” is describing this case, and it is the common case.
• Running vLLM with --tensor-parallel-size 2 on two RTX 3090s? NVLink is worth it. A published benchmark measured 715 output tok/s with NVLink vs 483 without — about +50%.
• Four GPUs? The gain collapses to about 10%, because a bridge pairs cards and half the traffic still crosses PCIe.
• The RTX 3090 is the last GeForce card with NVLink. The 4090 and 5090 do not have it and never will.
• Buy the bridge only after you have measured your slot spacing. Wrong spacing means the bridge physically will not fit.

Why Both Answers Are True

The disagreement is not about hardware. It is about how your inference engine splits a model.

Layer split (pipeline parallelism) — what Ollama and llama.cpp do by default. Layers 1–40 live on GPU 0, layers 41–80 on GPU 1. A token flows through GPU 0, one activation tensor crosses to GPU 1, and the token finishes. That is one small transfer per token, a few megabytes at most. Even slow PCIe handles it without breaking a sweat. The GPUs also take turns rather than working simultaneously, which costs you utilisation — but not bandwidth.

Tensor parallelism — what vLLM does with --tensor-parallel-size. Every layer is cut across both GPUs, both cards compute their half at the same time, and they must all-reduce their partial results at every single layer. For an 80-layer model that is 80 synchronised exchanges per token instead of one. Now the interconnect is in the critical path, and its speed shows up directly in tokens per second.

Same two cards. Same bridge. Completely different sensitivity — because one design touches the link once and the other touches it eighty times.

The Measured Numbers

The clearest public data on this comes from a benchmark run on RTX 3090s under vLLM, serving Qwen2.5-7B-Instruct-1M with 200 random prompts at a 32,768 max model length and a 220W per-card power limit:

The author’s own summary: NVLink improved tensor-parallel inference performance by about 50% on 2x 3090s and about 10% on 4x 3090s.

Read the four-GPU rows carefully, because they are the more interesting result. Four cards are slower per-stream than two in this test, and NVLink rescues far less of it. The reason is structural: an NVLink bridge joins GPUs in pairs, so with four cards half the all-reduce traffic still crosses PCIe, and the slowest path sets the pace. More GPUs is not more NVLink benefit — it is less.

Two caveats we will not paper over. This is a 7B model at a 220W power limit, not a 70B at stock power, and it is one published benchmark rather than a body of replicated results. The direction and the rough magnitude are trustworthy. The exact percentage on your model is not something anyone has measured for you.

Which Engine Are You Actually Running?

Most people reading this run the engine in the left column and do not need a bridge.

If you are unsure which you are on, you are on layer split. It is every desktop tool’s default. Our inference engine comparison covers what else changes when you move to a serving stack, and it is a real move — vLLM is a server, not an app.

The Honest Framing: NVLink Is Not the Upgrade

Here is the part that gets lost. Look at the table again: two 3090s without NVLink hit 483 tok/s; two 3090s with NVLink hit 715. But you only see either number because you switched to vLLM. If you are on Ollama today, the large win available to you is the engine change, not the bridge. The bridge is the second-order optimisation that follows it.

So the correct order of operations is:

1. Build the dual-3090 box. Run Ollama. Confirm it does what you want.
2. If you need concurrency or throughput, move to vLLM tensor parallel.
3. Then, if you are on exactly two cards, add the bridge.

Buying the bridge at step 1 buys nothing. This is why the “you do not need NVLink” advice is right for almost everyone who asks — and why it is incomplete for the small group already serving with vLLM.

What NVLink Actually Delivers on a 3090

NVIDIA’s own RTX 3090 specification page confirms NVLink support — listed as “NVIDIA NVLink (SLI-Ready): Yes” — but does not publish a bandwidth figure. The commonly repeated community measurement is roughly 112 GB/s aggregate bidirectional between two cards. We flag it as a community number because NVIDIA does not state it, and you should be suspicious of any page that presents it as a vendor specification.

The ratio matters more than the absolute value. NVLink is several times faster than the PCIe path the cards would otherwise use, and it also skips a hop through the CPU’s root complex. That is what an all-reduce at every layer notices.

The 3090 Is the End of the Line

NVIDIA removed NVLink from the consumer stack with Ada Lovelace. The RTX 4090 does not have it, the RTX 5090 does not have it, and NVIDIA’s stated reasoning was that the die area was better spent elsewhere now that PCIe Gen 5 provides adequate peer-to-peer bandwidth.

That has a buying consequence people miss. When you compare dual RTX 3090s against a single RTX 5090, NVLink is a capability in the 3090 column that no newer consumer card can restore at any price. It is a small point — it only pays off under vLLM tensor parallel — but it is a permanent one. Multi-GPU 4090 and 5090 builds are PCIe-only, forever.

For a used 3090 pair, work each card through our used RTX 3090 verification checklist before the return window closes:

EVGA GeForce RTX 3090 24GB — the last GeForce card that can take an NVLink bridge at all.

Buying a Bridge: Measure First

NVLink bridges are sold in fixed slot spacings — commonly 2-slot, 3-slot and 4-slot. They are rigid. If the spacing does not match the gap between your two cards’ connectors, the bridge will not seat, and there is no adapter.

Most dual-3090 consumer builds are not 2-slot. Two triple-slot-thick 3090s on a normal ATX board typically end up three or four slots apart. Before you order anything: count the slot positions from the first card’s NVLink connector to the second card’s, and buy exactly that spacing.

The bridge we can currently link is the 2-slot variant:

NVIDIA NVLink Bridge, 2-slot (900-53651-2500-000) — compatible with RTX 3090, A5000, A5500 and A6000.

This is 2-slot spacing, and it is probably not the one you need. It suits tightly packed dual-slot workstation cards. If your two 3090s sit three or four slots apart — which is the usual outcome with air-cooled triple-slot cards — this bridge will not fit, and we would rather tell you that than sell you a return. We do not currently have links for 3-slot or 4-slot bridges. Measure, then buy the spacing you measured, from wherever stocks it.

When to Skip the Whole Question

Skip NVLink entirely if any of these are true:

• You run Ollama, LM Studio or llama.cpp. The gain is noise.
• You have three or four GPUs. About 10% on the benchmark above, for real money and cabling.
• Your cards are 4090s or 5090s. Not supported.
• You have not yet built the machine. Build it, run it, find the actual bottleneck. Motherboard and CPU choice and PSU sizing will affect your build far more than a bridge will.

See Also

• Dual RTX 3090 vs RTX 5090 for Local LLMs — 48GB slow versus 32GB fast, at current prices
• Best 48GB VRAM Setup for Local LLMs — every route to 48GB, including single-card options that need no bridge
• How to Buy a Used RTX 3090 Safely — the checklist to run on each card
• Motherboard and CPU for a Multi-GPU LLM Rig — x8/x8 slot spacing, which also determines bridge spacing
• Ollama vs LM Studio vs llama.cpp vs omlx — the engine choice that decides whether NVLink matters
• GPU for Fine-Tuning vs Inference — where NVLink pays off most consistently

Sources

• NVIDIA GeForce RTX 3090 / 3090 Ti product specifications — NVLink (SLI-Ready) support confirmation, 24GB GDDR6X, 384-bit, 350W TGP
• Published vLLM benchmark on RTX 3090s (Qwen2.5-7B-Instruct-1M, 200 prompts, 32,768 max model length, 220W power limit) — the NVLink on/off throughput table
• Reporting on NVIDIA removing NVLink from the Ada Lovelace consumer generation, and the PCIe Gen 5 peer-to-peer rationale
• The ~112 GB/s bidirectional figure is a community measurement, not an NVIDIA published specification