A transformer model trains at large scale. Which accelerator does Google design for this kind of work?
Why C is the answer
TPU
TPUs are Google's accelerators, built for large-scale ML work.
In detail
- TPUs are Google's accelerators for this work.
- Compare price and performance for your model.
Use case
An illustrative scenario, not a real incident.
A team trains a large transformer on TPUs, which Google designed for this work.
See it in different ways
stateDiagram-v2 [*] --> FitsInMemory FitsInMemory --> DataParallelism: yes, but too slow FitsInMemory --> ModelParallelism: no DataParallelism --> Train ModelParallelism --> Train Train --> Tune: hyperparameter tuning Tune --> [*]
flowchart TD
Q{"What is the workload?"}
Q -->|"Large-scale training"| A0["TPU or GPU"]
Q -->|"Small jobs"| A1["CPU may be enough"]Key point: TPUs suit large-scale ML.
Why the other options are wrong
- A. CPU-only instance
- CPUs are slow here.
- B. A Cloud Storage class
- That is storage.
- D. Edge device
- Edge devices are for inference.