If your business is moving beyond running pre-trained models and into actually training or fine-tuning your own, the single hardest hardware question is how much GPU VRAM you need. Get it wrong on the low side and your training jobs simply will not start, throwing out-of-memory errors before the first step completes. Get it wrong on the high side and you have tied up scarce capital in memory you never touch. For a Nigerian company budgeting in naira against a moving exchange rate, that decision carries real financial weight.
This article is the training-specific companion to our general guide on how much VRAM you need in 2026. Where that piece covers the broad picture across gaming, creative work and inference, here we focus entirely on the business of training and fine-tuning AI models, and why this workload is in a memory class of its own.
Why training needs far more VRAM than inference
Running a model for inference, where you feed it a prompt and read the answer, only needs the model weights held in memory plus a little working space. Training is a different beast entirely. During training your GPU has to hold several large things at once, and they stack on top of each other.
- Model weights · the parameters themselves, the same as inference.
- Gradients · one value per parameter, computed on every backward pass, effectively doubling the weight footprint.
- Optimiser states · modern optimisers such as Adam keep two extra values per parameter, which can be the single largest consumer of memory.
- Activations · the intermediate outputs from every layer, kept in memory so gradients can be calculated, and these scale with batch size and sequence length.
A rough rule of thumb for full training is that you need somewhere around sixteen to twenty times the raw parameter size in bytes once weights, gradients and Adam optimiser states are all resident. That is why a model you can comfortably run for inference on a modest card may need several times more memory before you can train it at all.
Practical VRAM guidance by model size and method
The good news is that you rarely need to train from scratch. Most Nigerian businesses are fine-tuning an existing open model on their own data, and the method you choose changes the memory requirement dramatically. Here is a practical mapping.
- Small models and LoRA fine-tuning · 12 to 16GB is enough to fine-tune small models or apply low-rank adaptation to mid-sized ones, because LoRA only trains a tiny fraction of parameters.
- Comfortable 7B fine-tuning · 24GB gives you room to fine-tune a 7 billion parameter model with sensible batch sizes, the sweet spot for most first serious projects.
- 13B models and full fine-tunes · 48GB opens up larger models and full-parameter fine-tuning rather than just adapters.
- Anything larger · models in the tens of billions of parameters, or full training of 13B and up, push you into multi-GPU territory where memory is pooled across cards.
If you are deciding between adapter methods and full fine-tuning, our walkthrough on building a workstation for machine learning in Nigeria covers how those choices shape the rest of the build.
How batch size and sequence length change the maths
Two settings quietly drive a large share of your memory use, and both relate to activations. Batch size is how many training examples the GPU processes at once, and memory used by activations scales almost linearly with it. Sequence length, the number of tokens in each example, has an even steeper effect because attention costs grow faster than linearly with longer contexts.
The practical consequence is that the same model can fit or fail to fit depending purely on these knobs. If you hit an out-of-memory error, halving the batch size is often the fastest fix. This also means published VRAM figures are only ever a starting point. A 7B fine-tune at short sequence lengths behaves very differently from the same model trained on long documents, which matters for Nigerian businesses working with lengthy contracts, legal text or customer transcripts.
Techniques that stretch the VRAM you have
Before you spend on more memory, several well-established techniques let you train larger models on a given card. Knowing these is the difference between buying sensibly and over-buying.
- Gradient checkpointing · discards most activations during the forward pass and recomputes them during the backward pass, trading extra compute time for a large drop in memory use.
- Quantisation · loading model weights at lower precision, such as 8-bit or 4-bit, shrinks the resident footprint and is the basis of methods that let surprisingly large models fine-tune on consumer cards.
- Gradient accumulation · lets you simulate a large batch by summing several small ones, keeping memory low while preserving training quality.
- Mixed precision · doing most maths in 16-bit rather than 32-bit roughly halves activation and gradient memory and speeds things up on modern hardware.
These techniques are not free. Gradient checkpointing makes each step slower, and aggressive quantisation can affect final model quality on some tasks. But they often turn an impossible job into a practical one, which is why we recommend exhausting them before reaching for a bigger GPU.
Mapping requirements to real card tiers
Once you know your target, the hardware choice usually comes down to consumer versus workstation memory. The 24GB consumer tier covers a very large share of real business fine-tuning work and offers the best value per naira. The 48GB and above workstation tier exists for full fine-tunes, larger models and teams that want headroom to avoid constant memory wrangling. Beyond that, two cards pooled together serve the largest jobs.
If you want to understand how these tiers differ in practice, our breakdown of GPU tiers from entry to high end is a useful reference, and for teams committing to memory-heavy training, the workstation-class Blackwell cards are built precisely for this. As a rough naira guide, a capable 24GB consumer card and the supporting build sits well below a single 48GB workstation card, which in turn costs a fraction of a dual-card setup, so the steps between tiers are significant and worth justifying with real workloads.
How to estimate your own need before you buy
You do not have to guess. A sound approach is to pick the largest model you realistically expect to train in the next year, decide whether you will use adapters or full fine-tuning, and then add a margin for batch size and longer sequences. Where the numbers are tight, plan to lean on gradient checkpointing and quantisation rather than jumping a whole tier. The goal is to buy enough VRAM that memory never blocks the project you actually have, without paying for capacity that sits idle.
For Nigerian operators there are two local realities to fold in. First, power: training jobs run for hours or days under sustained load, so clean power and proper backup are not optional, a point we cover in optimising a PC for Nigerian power conditions. Second, the foreign exchange rate means imported GPUs are a meaningful capital outlay, so a card that lasts two or three project cycles is usually a better buy than the cheapest option that boxes you in within months.
Frequently Asked Questions
Can I train a model on a gaming GPU? Yes, for many real workloads. A 24GB consumer card handles LoRA and full fine-tuning of 7B-class models well, and with quantisation it can reach further. You only need workstation cards once you require their larger memory or run training nearly continuously.
Why does my model run fine but fail to train on the same card? Because training also stores gradients, optimiser states and activations on top of the weights, often needing several times the inference memory. Reducing batch size or enabling gradient checkpointing usually resolves the out-of-memory error.
Is it better to buy one 48GB card or two 24GB cards? For a single large model that must fit in one place, the 48GB card is simpler and avoids the overhead of splitting work. Two cards suit running several jobs in parallel or scaling beyond what one card holds, which our team can help you weigh.
The Bottom Line
Training and fine-tuning are memory-bound workloads, and VRAM is the constraint that most often decides whether a project gets off the ground. Match your purchase to the largest realistic job ahead, use the memory-saving techniques before jumping tiers, and treat the consumer-to-workstation gap as a deliberate decision rather than a default. Done well, you avoid both the frustration of being blocked and the waste of idle capital.
If you are sizing a training machine for your team, our configurator lets you match VRAM, compute and power to your actual workload, and you can contact us to talk through your specific models and budget with someone who builds these systems for Nigerian businesses every day.