Show HN: I made a calculator to show cost savings of serverless GPUs vs. AWS

Show HN: I made a calculator to show cost savings of serverless GPUs vs. AWS(paylesstoaws.com)

25 points by votick 4 years ago | 4 comments

archerx 4 years ago |

Or you can just buy a GPU instead of renting one out. I have a project where users generate designs for clothing and the mock ups get rendered in 3D on various different models, about 20 variations per design. Instead of paying a monthly bill to some GPU provider forever I instead bought a Nvidia Jetson Nano and set it up as a render server. It constantly looks at it's render queue and renders as needed. It cost me $150 to put it together and sometimes I forget it exists because it runs so well.

Sometimes it's cheaper and more practical to just buy the hardware.

I had a similar issue with a company I worked for before, we had an account on Azure and the machine was expensive but also slow and after 4 months we decided to buy a used server for $1000 and closed the Azure account and never looked back.

randomuser256 4 years ago | |

Do you mind sharing how you ‘deployed’ your own gpu? How did you host this?

archerx 4 years ago | | |

I researched and bought the Jetson Nano ( https://developer.nvidia.com/embedded/jetson-nano-developer-... ) because it would be just strong enough for my rendering tasks. I set it up in my home with the supplied Nvidia Linux distro (I stuck to the Nvidia provided distro because I assume it would have the Nvidia best driver support).

Once ready I opened up the render queue page in Chrome on it and as soon as a new job comes in it renders it. It get's it's jobs from a SSE connection. I use Three.js as the 3D library. It works very well now but in the beginning it had some hitches because while the GPU is strong the CPU on the Nano is very weak, so decompressing the models and the initial asset loading where not as fast on beefy modern x86-64 but nothing a few quick optimizations could not fix. The final hurdle was occasionally losing context but I just added a listener for those events and restarted the job if it failed. Another fun solution was checking to see if the image were actually rendered instead of just a transparent PNG by checking the file size of the outputted image, if the image is impossibly small it gets sent back to the job queue.

naikrovek 4 years ago |

buying GPUs will always be cheaper than renting, unless something changes dramatically. even with GPU prices being what they are today.

depending on usage, you could start saving money over rented GPU time in a single-digit number of weeks.