If you have shopped for a second GPU to speed up your AI work, you have probably hit a confusing wall of jargon: NVLink, PCIe, data parallelism, tensor parallelism, DDP, FSDP. The internet is full of strong opinions, and a lot of them will push you towards expensive hardware you may never need. The honest truth is simpler than the marketing suggests: for most people doing AI training and fine-tuning, ordinary PCIe is perfectly fine, and NVLink only earns its keep in specific situations. This guide explains those situations plainly so you can spend your naira where it actually counts.
We will assume you already know why two cards can be useful. If you are still at the planning stage, our step-by-step dual-GPU build guide walks through the full machine, and scaling from a single to a multi-GPU setup for a Nigerian business covers the commercial side. Here we focus on one question only: how the cards talk to each other, and when that conversation becomes a bottleneck.
Why use more than one GPU in the first place
There are two distinct reasons to add a second card, and they pull your hardware decisions in different directions. Knowing which one applies to you is the first step.
- More total VRAM. A single card can only hold a model that fits in its memory. Split a larger model across two cards and you suddenly have room for something that would never fit on one. This is model parallelism, and tensor parallelism is a common form of it.
- Faster training. Keep a copy of the same model on each card and feed each one different batches of data at the same time. You process more data per second, so training finishes sooner. This is data parallelism.
The first reason is about capacity. The second is about throughput. They place very different demands on how much the cards must talk to each other, and that is precisely what decides whether you need NVLink.
The interconnect question, in plain terms
During training, the GPUs sometimes need to exchange data with one another. How much they exchange, and how often, depends entirely on the strategy you chose above. The link they use to exchange that data is the interconnect, and you have two options.
- PCIe is the standard slot every GPU plugs into on the motherboard. It is universal, it works with every card, and it carries card-to-card traffic well enough for many workloads, though it is slower for direct GPU-to-GPU chatter.
- NVLink is NVIDIA's dedicated high-bandwidth bridge that connects two cards directly, bypassing the slower path. It is much faster than PCIe for card-to-card traffic, but it is only available on certain cards.
So the real decision is not "is NVLink better" — it always is, for raw card-to-card speed — but "does my workload move enough data between cards for that speed to matter". For a lot of real AI work, it simply does not.
When PCIe is perfectly fine
The most common way to train across two cards is data parallelism, usually implemented as Distributed Data Parallel, or DDP. Each card holds a complete copy of the model and works on its own slice of the data. The only time the cards must talk is to synchronise gradients, and that happens periodically rather than constantly. That is comparatively light cross-card traffic, and PCIe handles it comfortably.
This covers a huge share of practical use. If you are fine-tuning a model that already fits on one card and simply want it done faster, two cards on PCIe will serve you well. There is no reason to pay extra for NVLink bandwidth you will never saturate.
- You are fine-tuning a model that fits within a single card's VRAM.
- You are using DDP-style data parallelism to speed up training.
- Your cards do not support NVLink anyway, which is true of most modern consumer cards.
- You are running inference or smaller-scale experiments rather than constant heavy cross-card communication.
For these cases, the money you would have spent chasing NVLink is far better invested in more VRAM per card, faster storage, or a sturdier power setup. Our piece on fine-tuning small LLMs locally reflects this reality — most of that work never needs a fancy interconnect.
When NVLink genuinely helps
NVLink starts to matter when the cards must exchange a lot of data, constantly, at every step of training. This happens with heavy model or tensor parallelism, where a single model is split across both cards and the two halves must communicate continuously to complete each forward and backward pass.
The classic example is running or training a model too large to fit on one card. If you want to work with something like a 70-billion-parameter model, you have no choice but to split it, and the cards then talk relentlessly. Our guide to running Llama 70B locally in Nigeria goes deeper on that specific scenario. In this regime, a slow interconnect becomes the bottleneck, and NVLink's bandwidth meaningfully reduces it.
There is a second case: Fully Sharded Data Parallel, or FSDP. Instead of keeping a whole copy of the model on each card, FSDP shards the weights, gradients and optimiser states across the cards. This lets you train larger models than DDP allows, but it moves considerably more data between cards, so it benefits much more from a fast link.
- You are training or running a model too big for one card, using tensor or model parallelism.
- You are using FSDP, which shards model state across cards and moves heavy traffic.
- The cards communicate every single step rather than periodically.
- You have measured a real communication bottleneck and confirmed the interconnect is the cause.
DDP versus FSDP, simply put
The two strategies are the clearest way to understand the whole NVLink question, because each one tells you immediately how much your cards will chat.
- DDP keeps the whole model on each card and only syncs gradients now and then. Less cross-talk, PCIe-friendly. This is where most people start.
- FSDP splits the model across cards and moves far more data between them every step. More cross-talk, and it loves a fast interconnect like NVLink.
If those terms feel abstract, it helps to understand how a GPU moves data internally too. Our explainer on GPU memory bandwidth covers the same underlying idea — that moving data, not just crunching numbers, is often the real limit on performance.
The availability catch and a power warning
Here is the practical twist: you cannot always choose NVLink even if you want it. It is limited to specific cards. The RTX 3090 famously supports it, but many newer consumer cards dropped the connector entirely, leaving it largely to datacentre hardware. So if your strategy genuinely needs NVLink, that decision often dictates which used cards you hunt for in the first place. Our look at the used GPU market, RTX 30 versus 40 in Nigeria and our piece on whether a dual-GPU SLI or NVLink setup is worth it in 2026 both wrestle with exactly this trade-off.
Whichever path you take, do not forget power. Two GPUs under full training load draw serious wattage, and with NEPA being NEPA, an abrupt cut mid-run can corrupt a checkpoint or worse. Size your power supply with real headroom and put the whole rig behind a proper UPS so a flicker does not cost you hours of training. A multi-GPU machine that trips on every brownout is no machine at all.
Frequently Asked Questions
Do I need NVLink to use two GPUs for AI? No. Two cards work together over ordinary PCIe with no special bridge. NVLink only adds value when the cards must exchange large amounts of data constantly, as in heavy tensor parallelism or FSDP. For typical DDP data-parallel training and most fine-tuning, PCIe is entirely sufficient.
My cards do not support NVLink. Have I made a mistake? Almost certainly not. The vast majority of consumer cards never had it, and the workloads most people run — fine-tuning models that fit on one card, data-parallel training — never need it. You only lose out if you specifically plan to split very large models across cards every step.
What is the difference between DDP and FSDP in one line? DDP puts a full copy of the model on each card and syncs occasionally, so it is light on cross-card traffic and PCIe-friendly. FSDP shards the model across cards and moves far more data between them, so it benefits much more from a fast link like NVLink.
The One Thing to Remember
Start with PCIe. For fine-tuning and ordinary data-parallel training — which is what most people actually do — two cards on PCIe perform brilliantly, and the money saved is better spent elsewhere. Reach for NVLink only when you are splitting a genuinely large model across cards or using FSDP, where constant heavy cross-card traffic makes a fast interconnect worth every kobo. Choose the strategy first; let it tell you whether the interconnect matters.
Not sure which camp your project falls into? Build a balanced dual-GPU machine with our configurator, or get in touch and we will spec a rig matched to the exact training you intend to run — no overpaying for bandwidth you will never use.