online gpu rental services -- for AI
RunPod -- GPUS with pre installed configured MOdels -- and
cheap for running model per sec pulse --
https://salad.com/pricing -- very cheap --
Why salad so cheap:
We utilize unused, latent compute resources from gamers and e-sports arenas worldwide to power our network — think Airbnb for GPUs.
RunPod = real datacenters in US/EU - reliable, fast.
Salad = thousands of gamer PCs worldwide sleeping at night - they rent their GPU when not gaming. That's why it's cheap.
Salad good: Insanely cheap ($0.02/hr for RTX 5090), 191 countries
Salad bad: Distributed = unpredictable availability, Max 24GB VRAM per GPU, Not for low-latency workloads
- RunPod: Per-second billing
- Salad: You only pay for time hardware is available, they don't charge for cold boot. Hourly calculator, but billed per second.
- Both per hour, not per month, but you can keep running 720 hrs = monthly.
For batch jobs / non-critical - Yes, Salad is cheapest in world at RTX 3090 $0.09/hr, RTX 4090 $0.16/hr.
For interactive video work where you need GPU NOW - RunPod/Vast.ai better. Salad can be offline when you need it.
2. What happens if you take Salad vs RunPod?
Salad is Airbnb for gamer PCs.
- RunPod: Datacenter GPU, you rent same machine. You have Network Volume - install once, stays forever. You can keep it for months, stop/start.
- Salad: Every time you request, you get a different random gamer's PC somewhere in the world. You cannot keep it long time, max 24-48 hrs, and no Network Volume. You have to use Docker - you build your image with Python + models baked in, push to Salad. Each new PC downloads your Docker - 2-3 min load. If that gamer turns off PC, your job moves to another PC - interruption.
For your pattern of 20 times per day for 3 mins, Salad is actually okay because it's cheap and stateless, but for long 1-hour video work, RunPod is more stable.
Install every time?
- RunPod Pods: No, with Network Volume install once.
- Salad & RunPod Serverless: You bake models into Docker image once, then every launch loads from Docker - 1-2 min, no manual pip install each time.
Comments
Post a Comment