Setting up our own AI server

Discussion in 'Hardware' started by MarkBrown, Oct 17, 2024.

  1. MarkBrown

    MarkBrown

    thinking about going this route Llama

    been researching hardware requirements and seems bigger is better as usual.
    https://llamaimodel.com/requirements-3-2/

    has anyone else looked into this and the different models specifically for coding trading systems?

    i have been using chat gpt with good success tackling some heavy math experiments that are not even possible in matlab after years of messing with that. i love the chat gpt but its very forgetful of historical corrections and the more you use it the slower and more mistakes it makes.

    m

    for sure the rtx 4090 is on the shopping list and the nvme drive. but as far as processors torn between amd runs hot and intel probably not as fast. i know dell makes a 30k ai machine and probably hp - wonder if they will sell these eventually on the cray.
     
    Last edited: Oct 17, 2024
  2. Businessman

    Businessman

    LLMs == curve fitting
     
    EdgeHunter likes this.
  3. QuantVPS

    QuantVPS Sponsor

    You might find this interesting: Jane Street Real-Time Market Data Forecasting Competition. It could offer some valuable insights. Kaggle has a ton of great resources like datasets and models.

    If you’re considering running an LLM locally, I’d recommend checking out LM Studio—it’s a great tool. Pairing it with a MacBook Pro M3 could give you excellent performance, especially with Apple’s chips being so strong with these models. It is nearly pointless to build a PC. You can also explore AnyModel to compare public models easily.

    However, if you’re thinking about training your own models, keep in mind that a single RTX 4090 might not be enough, depending on the dataset size and complexity of the model. It is likely best to outsource this to a provider that can handle this type of compute (such as AWS or QuantVPS).
     
    MarkBrown likes this.
  4. Hi Mark,
    I have researched this too but I just don't think the open source local models are there yet compared to the big players.
    My current AI assisted programming is $40 a month for Cursor, chatgpt o1-preview and then Claude Sonnet free tier if needed. I think chatGPT is behind right now after the 10/22 update to Claude Sonnet and not really worth it.
    Building a massive rtx 4090 machine would be fun though. You could even have multiple rtx 4090 if you want to spend the money.

    Building a cheaper 3060 machine might be better value though just to play around for the future.

    Right now though I just don't see a reason to not use the most powerful model and that to me is easily Claude Sonnet. Before 10/22 that was o1-preview. Next week it might be something else.