Viable paid options?

[carmine prawn]

Colab, paperspace and kaggle are so shit and slow. Is there any viable option to train my voice model that's not free/freemium? You mostly recommend free options, but GPU training for a model is not that expensive you can get away with <$4 per hour which could get you really high speed and training on A100-H100's. Any ideas? IBM? AWS SageMaker? Please let me know

[median fiber]

Vast.ai

[valid geyser]

#🔊│ai-development message
though haven't seen anyone trying to deploy on AWS

for normal RVC training, even consumer gpus with 6-8 gb vram are viable

[valid geyser]

also in order to be faster, use more batch size (might have worse generalization but more "stable") and/or shorter dataset

[carmine prawn]

I have a 3080, took 8 or so hours with full gpu utilization to train 500 epochs...

20 minute dataset

in Applio

Colab on A100 gpu took around 1 hour

[valid geyser]

yeah that's normal, batch size 4 ig?

[carmine prawn]

yeah so it means Ihaveto do it overnight, hell nah

i'd rather pay a couple dollars

don't remember 4 or 8

whatever the guide said

[valid geyser]

iirc training ~15 min dataset on colab took about 5 h overall for 400 epochs with batch size 8

[carmine prawn]

Yeah definitely not worth it

Especially with the risk of it shutting down, even with the afk javascript injection

I mean, with all due respect, I asked for specifically a paid service that allows me to use high performance GPUs, I don't see why you're trying to convince me of using my GPU

Just leads to unnecessary arguments, don't waste your time helping people with something they didn't ask

[valid geyser]

i dont see any point of not using ur 3080 unless offloading it for ur other game/workloads

[carmine prawn]

You think my 3080 is just laying around doormant when im on my pc or what?

Why would I have a 3080 then?

[valid geyser]

also not many people have/afford gpu as powerful as 3080 or higher

[carmine prawn]

So? What does that have to do with my request here exactly

I don't understand your thought process here

Some people can't afford GPU's like 3080, therefore I should not train my model on higher end gpu's on the cloud?

Did you just enter this discussion to moral police? 🤣

[valid geyser]

tf u mean "moral police"?

well do u have some use cases that require more than 12 gb vram?

[carmine prawn]

Nope, if I did, then I would go for a 16gb or higher GPU

[valid geyser]

also the A100/H100 pricing alone is about five-digit bucks, so dont expect more affordable renting on it

[carmine prawn]

No it's not? It's $2 per hour

And it increases my speed by factor of 20-30x

The electricity generated by having a 3080 run on 100% almost max temp and power consumption for 6-8 hours is probably more expensive than this

And it gives me plenty of time to play around with different batch sizes, pre-trainers and many more opptions since I can do multiple runs in hours.

Instead of once per day

[valid geyser]

with that overkill vram, you could try training several models at the same time