Do standard web hosts support GPU-accelerated AI workloads?

Question

Trying to figure out if I can just use a standard web host for some light image processing stuff. Ive been reading that shared hosting is basically a no-go for anything GPU-heavy but then I saw a couple threads claiming some VPS providers let you pass through hardware. Im based in the UK and only have like 30 quid a month to burn, so a massive dedicated cloud GPU instance is totally out of my budget. I just need to run some basic neural network inference for a personal side project by next month. Do I actually need a specialized GPU cloud or is there some hack to get this running on a normal cheap host?

uzjgfhfitl · Accepted Answer

Honestly, forget shared hosting. I tried running inference on a Linode Dedicated CPU 4GB instance once and it was just painful. You really need actual GPU access for these models.

tfikvxnfqs · Answer

Honestly, shared hosting will just crash your app. I tried running inference on a cheap VPS last summer and it was a disaster. I ended up finding some affordable GPU cloud options that saved my project! You should really look at these instead:

RunPod RTX 3090 24GB VRAM

Vast.ai RTX 4090 24GB VRAM It is so much faster for way cheaper than a dedicated server. Total game changer for my workflow!

RobertHoish · Answer

^ This. Also, dont bother with standard web hosts. They block anything that burns CPU cycles for too long, and GPU pass-through is basically non-existent on those plans. I tried getting cute with it before and my account got suspended within 20 mins for resource abuse. Not worth the headache. Since you're on a strict 30 quid budget, standard VPS providers will just disappoint you. Youre way better off renting hourly on decentralized GPU marketplaces. Here are a few ways to stretch that budget:

Lambda Labs NVIDIA A10 24GB VRAM - super stable but you have to catch them when they have stock

Paperspace Gradient NVIDIA Quadro P4000 8GB VRAM - great for smaller inference tasks without breaking the bank Just stick to these specialized clouds and you wont have any issues with hardware access.