Ask HN: What desktop to buy for home based AI/ML in 2023 | Hacker News Reader

Ask HN: What desktop to buy for home based AI/ML in 2023

5 pointsImageXav8 comments

This year I started running some AI/ML experiments at home using Google Colab. However, I want to start using my own compute for some of these.I was thinking of buying the following PC: https://www.palicomp.co.uk/amd-velocity-fac10 and upgrading to a 3080 and i7 as I do a lot of image analysis and it would be helpful to be able to parallelism some operations.

What does HN think of this and these specs? Any recommendations? I'm trying to keep my budget under £2000 unless going a bit over that would allow for a real increase in processing power.

View on HN

I would recommend a card with higher RAM, may be a 3090ti coupled with an AMD Ryzen 9 processor. Recently, I got one for around 2000 USD. https://store.nvidia.com/en-gb/geforce/store/gpu/?page=1&lim...

I was thinking of going for a 3090 but wasn't sure whether it was worth it due to the price still being quite high, it would certainly blow my budget out of the water. What kind of models are you planning on using it for?

AMD Ryzen 9 processors definitely seem worth looking into though, thanks.

If you are planning to do serious DL projects with images or anything multi-modal single 12GB card might not get you very far and you could be severely limited during training.

I would pick AMD Ryzen CPU and any Nvidia GPU with more ram.

If budget allows, I would pick RTX3090, if budget is too tight, look at RTX3080 or RTX3080Ti 12 Gb vram models. 10 gb vram models can be too small in the future or if you plan to use it for image processing.

I also am working mainly with image processing and segmentation, we mainly use RTX3080 and RTX3080Ti GPUs.

It should also be cheaper to pick components and build PC yourself.

I strongly recommend using colab or a crappy gpu for experimentation and just paying for compute when you want to run jobs that need it. Most tinkering doesn't need any special gpu, so whatever you buy is going to mostly sit idle. Paying for cloud GPUs obviously costs more per hour, but you then only pay for what you need and can scale however you want

The problem with Colab is that I tend to run into two issues:

1. Disconnect during training 2. Not enough storage

I tends to work on somewhat mid/large models and TB+ datasets so I run into these issues frequently enough that it's become a hurdle I want to avoid going forward. I suspect this might make online compute expensive too though I've not looked at pricing plans. Where would you recommend buying personal cloud compute?

Are you planning on making money doing this? Invest in a 7950X and all the RAM you can and either a:

4090

3090

3090 Ti

A30

A40

A100

H100

Not at the moment. I have a few hypotheses that I'd I'd like to test and I've reached the bounds of what Colab can do.

Reply on news.ycombinator.com