Hello, I’m new to AI and ML. I was exploring options to rent GPUs but they are so freaking expensive, I was wondering if I can setup my own ML environment on a PC? Can anyone guide me on this topic?
I’m glad that you asked. Few years back I was also in the same boat. I did a lot of research and pros and cons of renting GPU vs Owning my own machine for ML and AI work. I’m totally not in favour of renting a GPU and the reason is simple, cost. When you buy your own system it remains with you for almost like decades.
I’ll share my story how I setup my first GPU system for training AI model.
Few years back GeForce RTX 3090 was the latest GPU and also recommended for the ML. I bought a customized PC with watercooled RTX 3090 GPU. It came with Windows 11 so I installed Ubuntu as well.
In Ubuntu I trained my model and it worked. Not only it worked I also used the AI model generated on this PC in my live app which helped me a lot to increase my earnings.
I also used the system to mine crypto and made some money but I stopped using it after energy prices went up and Ethereum changed their algo.
There is one problem though. If you are using Ubuntu to train your model then chances are that you may end up fixing NVIDIA driver issues. NVIDIA and Linux don’t go well together. NVIDIA tech support is totally rubbish so don’t expect them to provide you any support. It is just mess. After months of struggle I did manage to get NVIDIA driver installed on my Ubuntu which was wiped out when I upgraded Ubuntu and since then I’m not able to fix it at all.
I don’t use GPU now a days so can’t provide you any updates on Linux vs NVIDIA fuckups. May be things have improved.
So to sum up just buy or get a custom PC build around latest NVIDIA GPU and that is all you need to get you started.
Oh WOW! thanks for detailed replies. I heard that water cooled PC are hard to maintain, what is your personal experience on that?
Yes, it is messy to change cooling liquid. My PC builder didn’t install a simple knob to drain out liquid so I’ll have tough time draining it out. My machine is very heavy so It will take some hard work to replace cooling liquid.
Luckily I never had to change the liquid so can’t comment on the practicality of that. I hope I’ll not have any trouble in changing liquid when it requires in the future.
Another question, should I build PC myself or get it custom build from PC builders?
If you have time, know how and most importantly some space in your house then you can build it yourself. You can buy all the required components from online stores such as Shop Moksha and other similar stores. I decided to buy because I didn’t have space or time to do it myself. Building a machine shouldn’t be that difficult.
Linux is great when everything works. Once something is broken specially the drivers then you waste rest of your life fixing it. For me, personally, Wi-fi drivers have been PITA to get working. Every new machine on which I’ve installed linux has always given me trouble with Wi-Fi.
When you get you machine built use Ethernet for Internet connection.
You don’t have to use Linux for ML. you can use Windows, CUDA works on Windows as well.