|
|
|
|
|
by 013a
2879 days ago
|
|
Pricing and capabilities have probably changed over the years. The cheapest GPU spot instance you can get on AWS today is a p2.xlarge, which will run you $0.45/hr. And also, p2-class instances run Tesla K80 GPUs, which are Kepler cores, an architecture running on 6 years old now (think GTX 600/700). Fine for 1080p60 most games, but I doubt many modern titles would hit the 60fps mark. Google Cloud can allow you to get the cost lower (closer to ~$0.15/hr) by cost optimizing the instance CPU/Memory and attaching a GPU to it. But you've got three problems with both of these setups: - Spot on AWS/Preemptible on GCP will kill your instance with about 2 minutes of warning. I'm not sure what AWS's policy is on how often this happens, but on GCP, its guaranteed to happen at least once every 24 hours. So that could be annoying if you're in the middle of an intense game of Overwatch. - Storage costs. You're paying per hour for your instance, so you don't want to have to download all your games every time you spin it up. But, then you shift the cost to the SSD, which is expensive and must be maintained 24/7/365 for that quick startup time. You could alternatively store it in S3/GCS then hot-load it when the instance starts, which would be fast due to the fiber interconnect all these datacenters are wired with. So, plan for this bullet to add at least $10/mo/TB of storage. - Network costs. A 1080p60 stream is, conservatively, 10Mbps? Add on an extra ~$0.25 per hour you want to game, just in bandwidth. |
|