I'm building a short-lived GPU lab for experiments—what's the cheapest way?

If you want the cheapest short-lived GPU lab, the answer is usually don’t buy a GPU box unless you’ll use it a lot. For experiments, the lowest-cost path is typically one of these:

Cheapest options, in order

1) Cloud GPU instances with preemptible/spot pricing

Best when you need GPUs only occasionally.

• Pros: no upfront hardware cost, quick to start/stop, pay only while running
• Cons: spot instances can be interrupted, costs can add up if left on
• Good for: training runs, prototyping, short experiments

Tips to keep it cheap:

• Use spot/preemptible/interruptible instances
• Pick the smallest GPU that works for your workload
• Shut down automatically when idle
• Store data on cheap object storage, not on the GPU instance
• Use a single GPU unless you truly need more

2) Buy a used consumer GPU and build a bare-minimum local machine

Best if you’ll run experiments repeatedly over weeks/months.

• Pros: low long-term cost, no hourly charges
• Cons: upfront spend, power/noise/maintenance
• Good for: stable dev environment, repeated iteration, offline work

Cheapest local setup:

• Used RTX 3060 12GB / 3070 / 3080 / 3090 depending on memory needs
• Used office PC or basic used workstation
• 32–64 GB RAM
• Decent PSU
• SSD only, no fancy extras

Important: For ML/graphics experiments, VRAM matters more than raw speed. A cheaper card with more VRAM can be more useful than a faster card with too little.

3) Rent a dedicated GPU box for a month

A middle ground.

• Pros: predictable, simpler than spot instances
• Cons: usually more expensive than spot cloud
• Good for: a defined short project where interruptions are painful

Rule of thumb

• A few hours to a few days: cloud spot/preemptible is usually cheapest
• A few weeks to a few months with frequent use: used local GPU may be cheaper
• Need maximum convenience: dedicated rental or on-demand cloud

How to make either option cheaper

• Use Linux rather than Windows
• Avoid overbuying CPU/RAM; GPU and VRAM usually matter most
• Use containers so you can recreate the environment easily
• Stop instances when not actively running
• Compress datasets and keep them outside the expensive compute node
• If cloud: choose regions with lower GPU pricing

If you want the absolute cheapest practical setup

• Cloud spot GPU for experiments
• Keep code/data in cheap storage
• Use a small instance and run jobs in batches
• Only move to local hardware if usage becomes frequent

If you tell me:

1. your workload type,
2. how many days/weeks you need it, and
3. your budget,
   I can suggest the cheapest specific setup.