I am planning to build a new PC and I want 2 things from used market
But I want suggestions from you guys as i might be missing something
Which processor and GPU for heavy AI and light gaming? My current budget is 50-55k
In general if you don’t know what you will do with “AI” - get maximum VRAM you can get, and I believe 3060 is indeed the best option.
If you are specific about exactly what AI workload you will be running, then someone can give you more specific advice. Sometimes getting a better GPU with lower VRAM can be better if you are certain the VRAM is enough for your target workload, because it will be faster.
3060 won’t be able handle “heavy” AI but it is bare minimum as most of the current models won’t fit in anything less than 12 GB. For training LLMs anyways single consumer grade GPU with humble i5 won’t be sufficient.
Interestingly of you’ll go for GPUs with bigger capacity upfront cost would be make you realize unless you are running constant inference or training small models a setup with Google Collab with some free tier and some paid credits would very very well.
This all was written keeping in mind you want to do heavy AI workload. For gaming this setup would run fine.
Going with RTX 4060 Ti 16GB is better than RTX3060 12GB for your use case.
Just post WTB here in TE for this GPU and buy.
Depending upon what exactly you want to do even with respect to inference and how many more GPUs you can add; you can go with dual 3060 12GB which will yield you ~24GB vRAM; and in the use market, you may be able to get 2 of those for just a little more than 25K, if you find the right deal.
It would be enough to run a handful of models at a decent rate, alternatively, you can also run them parallely to have 2 smaller models, depending upon your requirement.
Just make sure that you get a board that supports lane bifurcation/division.
If you want to get only a single card, depending upon the specific type of models you want to run and the software stack you use, you can probably get the B50 Pro/B580 [ under 30K ] and n the other hand, you’ll have 4060TI & 9060 16GB.
Though remember, you may exceed a bit more than the allocated budget for this.
The best way to go around with this would be to have an idea of what you want to run & your expectations before you jump into the hardware.
What kind of AI workloads are we talking about? Different workloads demand different vram, inferencing llm’s locally will demand huge memory, training basic CNN’s doesn’t. It will mainly depend on the things you’ll be working on. If you want to speed up training and inferencing, newer gen GPU’s would be better because of latest tensor cores, so it is all subjective unless we know what do you plan on using the new gpu for.
I am not sure currently, but I am planning to shift my carrier to AI field and mostly everything I will try and then stick to 1 thing eventually
If you’re not sure about it, don’t start pouring money into it.
Try getting inexpensive VPS with enough GPU compute for whenever you have the usecase.
Alternate answer, from your post I am assuming you have not dabbled far into AI yet. You can get an Nvidia GPU that fits your budget to experiment with toy learning problems but for “heavy AI” GPU vRAM matters - the more, the better. Even to run a relatively small LLM (~1B params) with good token output you need significant VRAM. I would advise start with free services like google colab, build your understanding then move onto cloud providers like jarvislabs (indian), vast.ai, runpod. Post that you can evaluate if you need a local deep learning setup “What is chiefly needed is skill rather than machinery” - Wright Brothers
Try using free colab to get your feet wet, then try the rented gpu’s - that shall give you an idea about what it is that inerests you and come to a conclusion for hardware.