The big idea: Training a model is enormous, one-off work: the same calculation over millions of examples, many times over.
Using a trained model is small, repeated work: one calculation per question.
They need completely different hardware.
Training
- Runs for hours, days or weeks
- Needs massive parallel processing — GPUs, TPUs, HPC
- Needs the whole dataset in fast storage
- Usually done once, then occasionally repeated
- Rented rather than bought, for most organisations
Using the model
- Runs in milliseconds, per request
- Needs far less power — sometimes a phone is enough
- Needs only the finished model, not the data
- Happens constantly, wherever the answer is needed
- Often on an edge device, for speed and privacy
Free preview
This is the free notes preview
You're reading the free notes. Aimnova Pro unlocks the full study experience — and you can try it with your first topic free to keep:
- FlashcardsLock in vocabulary and key terms with spaced repetition.
- Practice questionsAnswer exam-style questions and get instant AI marking.
- Mock exams & past-paper vaultSit full mocks and see exactly how examiners award marks.
- Personalised study planA daily plan built around your exam date and weak areas.
General purpose
- Standard laptop — fine for learning, small datasets, using a trained model
- GPU — thousands of simple cores; the standard choice for training
- Cloud platform — rent that hardware by the hour instead of buying it
Built for the job
- TPU — a chip designed specifically for machine-learning calculations
- ASIC — a chip built for one task only; fastest and most efficient, but fixed forever
- FPGA — a chip that can be reconfigured after manufacture; between a GPU and an ASIC
Where it runs
- Edge device — small hardware in a camera, phone or vehicle
- HPC centre — many machines together, for the largest training jobs
- Cloud and HPC for training; edge for using the result
ASIC and FPGA, in one line: An ASIC is built for one task and can never do another — fastest and most efficient, but useless if the task changes.
An FPGA can be reconfigured, so it is slower than an ASIC but survives a change of plan.
Parallel cores: the property GPUs and TPUs have in huge numbers, and the reason they suit training.
Interactive diagram
Explore the labelled diagram, charts and maps for this topic in full study mode.
Feeling unprepared for exams?
Get a clear study plan, practice with real questions, and know exactly where you stand before exam day. No more guessing.
A student experimenting — a laptop
Small datasets and small models, with no deadline.
A laptop is enough, and buying anything more would be waste. A cloud GPU can be rented for an hour if one experiment needs it.
A company training a large model — cloud GPUs
Weeks of heavy parallel work, needed occasionally rather than constantly.
Renting means paying only for the hours used, and scaling to more machines when a deadline demands it — with no hardware sitting idle afterwards.
A camera recognising faces — an edge device
The model is already trained; it just needs running, immediately, on the spot.
Processing locally means an answer in milliseconds, no video crossing the network, and it keeps working if the connection drops.
A phone manufacturer shipping millions — an ASIC
The same fixed task in every device, forever.
A chip built for exactly that job is the fastest and the most power-efficient, and the huge design cost is spread across millions of units.
Storage and scalability count too: Training data can run to terabytes, and it must be readable fast enough to keep the processors busy — slow storage wastes expensive hardware.
Scalability is about adding machines when the job grows, which is where cloud and HPC earn their place.
How this is tested — you must say whether the scenario is TRAINING or USING, then choose from that. It comes up two ways:
Paper 1 Section A
- Describe hardware for a stated scenario, 3-5 marks
- Distinguish two kinds of hardware
- Explain why one suits a described job
Paper 1 Section B — case study
- Recommend hardware for the case study's work
- Justify it from their scale and budget
The classic trap: Recommending powerful hardware for a scenario that only uses a trained model. Once training is done, a phone may be enough — and a GPU cluster would be an expensive way to answer one question at a time.
A company will train a large image-recognition model once, then run it on thousands of security cameras. Describe suitable hardware for each stage.
Model answer plan
See the answer broken into clear points, with guidance on how each point earns marks.