Training needs GPUs for a few hours a week. Which compute fits?
Why B is the answer
A compute cluster that scales down to zero when idle
Clusters scale to zero nodes when idle, so you pay only while training.
In detail
- A compute cluster scales to zero, so you pay only while it trains.
- Set a sensible idle time before it scales down.
Use case
An illustrative scenario, not a real incident.
A team trains a few hours a week. A compute cluster scales to zero when idle, so idle GPUs cost nothing.
See it in different ways
mindmap
root((Reproducible run))
Code
Git
Data
Versioned data assets
Environment
Registered environment
Compute
Cluster that scales to zero
Results
MLflowstateDiagram-v2 state "Zero nodes" as s0 state "Nodes running" as s1 [*] --> s0 s0 --> s1: job submitted s1 --> s0: idle
Key point: Clusters that scale to zero save money.
Why the other options are wrong
- A. A compute instance left running all week
- You pay for all the idle hours.
- C. A dedicated virtual machine left running all week
- You pay for all the idle hours.
- D. A managed online endpoint sized for production traffic
- An endpoint serves predictions, not training.