It started here on Sanyam’s mega gig post.
I am seeing very attractive posts from ML experts on how to make a massive rig that gives you the power to do stuff which only a few people can. It can be either winning Kaggle competitions, train models and push them to Huggingface or try experiments for their research papers.
I am doing neither of them. When I started learning deep learning 4 years back, my old laptop shitted on me. So I got a desktop setup considered decently powerful at the time with 8GB 1080, i5 and 16GB RAM. Recently I considered upgrading but it comes with cascading cost implication which is difficult to justify.
This post is not just about making that ultimate GPU rig. It is about selecting all the components such that you can enjoy working. I have been curating these for a while and want to share them.
There is not much content about how to make a home set up in India. Buying from India comes with its own challenges. And, how to make a setup for poor humans? I am going to take a stab at this.
Pro tip: Request your company for a WFH setup bonus. I got 35k when I joined my new company which eased me into spending on keyboard and chair.
Price - 39k
Sooner or later you will realise you are sitting all day on a chair breaking your back.
You cannot work standing because the height of the desk does not work out for you. The height needed while sitting on a chair is a function of chair height, length of hands, length of legs. You can get it working only if you can adjust the height. So you have 2 ways. Either find out the height needed while sitting/standing and get something done customly via a carpenter or buy the above one.
It’s definitely overpriced but why not? Treat yourself something you can rave about :P
Price - 20k
Don’t buy a chair with lumbar support and see your back pain rise eventually to the point it stops you from working more than a few hours.
A good chair is extremely critical.
Price - 27k
At times you want to code with split-screen. I have 27” but it’s not working out. Hence I suggest this one which is 34”. I really want the curved one but it’s unjustifiably costly - 55k.
Price = 60k for 3060 8GB
You can go with RTX 3060 8GB/Ti or 12GB. Many suggest RTX3070 too. There are other websites as well where you can buy them cheaper. You will be able to train BERT size models pretty fast with RTX 3060 8GB. I am on 1080 8GB which is 30% slower than 3060.
Check this thread for full-fledged research.
Price - 9k + 2.5k customs (26% import duty)
You won’t know how a good keyboard feels like until you have one.
Don’t try over researching this. I did it for you here.
This works wired with my desktop and on Bluetooth with my m1 pro.
Price - 15k
Quality is top-notch and the battery is insane. Working out great for me.
Note: This will require buying a motherboard with Bluetooth. Works with Macbook m1 pro.
Price - 4k
You will need a webcam if you are on a desktop.
You got to have this. The buttons increase productivity for sure! The scrolling is amazing. The battery is rechargeable. Can connect with laptop and desktop.
Total setup cost
Currently, I do not have the table due to space constraints at home but I will be having the rest of them soon. There are other things worth considering such as laptop stand and background lighting.
Total rig cost
= 39 + 20 + 27 + 60 + 11.5 + 15 + 4 + 60(other desktop stuff - RAM, motherboard, SSD)
= 2.4 lakhs
If you are not a serious GPU user, instead of burning it all on one RTX 3090, try to buy the little things which can make the experience of working more enjoyable and comfortable.
Except for the desktop-cpu, you can use all other items with your company laptop. This reduces your expenditure by 1.2 lakhs. You can also use the laptop as primary device and use the desktop as a server by starting jupyter on it.
Once you spend this, you don’t need to worry for at least 3 years. Only the electronics will deteriorate and the rest will have a long life. This translates to ~30k/year. I think it’s totally worth it.
What do you think?
Come join Maxpool - A Data Science community to discuss real ML problems!