GPU vs CPU: What Is Faster for Training Your Own NAM Model?
Capturing your own amp is the most exciting part of NAM. The recording takes minutes; the training can take minutes or hours depending on your hardware. Here is what actually makes the difference.
Create NAM Hub team · October 2, 2026 · 6 min read
How NAM training works, in one paragraph
You play a standard training signal (a special audio file provided by the NAM project) through your amp or pedal and record the result — this is called reamping. The trainer then teaches a neural network to turn the dry signal into the recorded one. Training repeats the same calculation many times over the audio ("epochs"), and that is the part that needs computing power.
The short answer
- NVIDIA GPU (CUDA): the fastest common option. A standard model with default settings usually finishes in minutes to roughly half an hour, depending on the card.
- Free cloud GPU (e.g. Google Colab): no special hardware needed — you upload your files and train in the browser. Speed depends on the GPU you are given and on availability.
- Apple Silicon Mac: can train locally; usually slower than a dedicated NVIDIA card but much faster than an older CPU.
- CPU only: it works, but it is slow — often hours for a standard model.
What affects training time
- Model size. Standard models take longest; Lite, Feather and Nano train faster and run lighter, at a small cost in accuracy.
- Number of epochs. More epochs usually mean a slightly better fit and a longer wait. Default settings are a sensible starting point.
- Hardware. GPU memory and generation matter more than the CPU when training on a GPU.
- Other programs. Close games, browsers with many tabs and video editors while training locally.
Local GPU vs cloud: pros and cons
| Local NVIDIA GPU | Free cloud notebook | CPU only | |
|---|---|---|---|
| Speed | Fast | Varies | Slow |
| Cost | Hardware you may already own | Free (with limits) | Free |
| Setup | Install drivers and trainer | Browser only | Install trainer |
| Best for | Many captures | First captures | Occasional, patient users |
Tips to get a good model faster
- Record cleanly. A good recording matters more than training time: correct levels, no clipping, no extra noise, and the reamped file perfectly aligned with the training signal.
- Use the standard training file and 48 kHz, exactly as the NAM project recommends.
- Start with default settings. Change the architecture or epochs only if you have a reason.
- Check the result by ear, not only by ESR. Compare the model with the real amp at the same volume.
- Batch your work. Record several settings in one session, then train them one after another.
Do you need a GPU to play NAM?
No. Playing through a model runs on the CPU in real time; the graphics card is only used for training. For smooth playback, see our latency guide.
Want to try capturing? Read Create your own NAM step by step.
Keep reading
How to Read a NAM Model Description: Gain, Cab and Mic
What to check in a NAM file listing before you download: capture type, amp settings, cab and mic, input level, model size and ESR.
Beginners5 Common Beginner Mistakes the First Time You Run NAM
Missing IR, two cabinets in a row, wrong input level, effects in the wrong order and audio settings — the five most common NAM beginner mistakes and how to fix them.
BasicsWhat Is an IR (Cab Impulse Response) and When Does NAM Need One?
Impulse responses explained for NAM users: what an IR is, when to use one, how to choose speakers and mics, and how to avoid the double-cab mistake.