OpenAI ChatGPT Server

[AllGoodNamesRGone]

Budget (including currency): $2500

Country: USA

Games, programs or workloads that it will be used for: OpenAI

Other details (existing parts lists, whether any peripherals are needed, what you're upgrading from, when you're going to buy, what resolution and refresh rate you want to play at, etc): 

I asked the chat server what it would take to build a openai chat-gpt, and below is the response:  

 

"To create a ChatGPT model within Ubuntu 22.04, you will need to have a machine with a CUDA-enabled GPU, as well as the following software installed:

Python 3.7 or later
CUDA 11.0 or later
cuDNN 8.0 or later
The Hugging Face transformers library
The torch library
Once you have all the necessary software installed, you can use the Hugging Face transformers library to fine-tune a pre-trained ChatGPT model on a dataset of your choice.
You will also need to download the pre-trained weights for GPT-2 or GPT-3 models from Hugging Face model hub or OpenAI API.
You can use the provided code examples and documentation from the transformers library to train and fine-tune the model to your specific use case.
It's worth noting that GPT-3 models are quite large and resource-intensive, so it might take a powerful machine with a large amount of memory and storage to train them.

Please note that creating a GPT model requires a good knowledge of deep learning and computer science, if you are new to this field I recommend to start with a simple project or to follow a tutorial that can guide you through the process."

 

From that response, I started to look at GPU's with CUDA-enabled, however quick to think I have no idea what I am looking at much less what to look for on price and specs.  At the end of this month is the last semester of college, studying python, how cool would it be to have figure this out by then.  

 

JD

[Takumidesh]

doesn't GPT-3 need like 1 terabyte of vram?

[Zando_]

4 minutes ago, Takumidesh said:

doesn't GPT-3 need like 1 terabyte of vram?

From some googling, looks like ~300-350GB VRAM for the smallest version of it, so yeah closer to 1TB for the largest sounds about right. I don't think you can get a single 48GB VRAM card for $2500, let alone the multiple OP would need. 

[AllGoodNamesRGone]

This is my reasoning behind building my own system, when to the website, and typed in the question and received the response  "An error occurred. If this issue persists please contact us through our help center at help.openai.com."  , high volume of people are trying to access the system, thus making it unavailable to the average user.  

[AllGoodNamesRGone]

https://www.projectpro.io/article/gpus-for-machine-learning/677#:~:text=The GIGABYTE GeForce RTX 3080,than with a different GPU. good page that talks about the GPU's and deep learning

[AllGoodNamesRGone]

HP L40549-001 8GB Nvidia Quadro RTX 4000 Turing Graphics Card - GDDR6 - 256-Bit - 2304 Cuda Cores - PCIe 3.0 X16 - 1005 MHz - Multi Monitor (Renewed) with tax was $730 on amazon.  

[Takumidesh]

5 minutes ago, AllGoodNamesRGone said:

HP L40549-001 8GB Nvidia Quadro RTX 4000 Turing Graphics Card - GDDR6 - 256-Bit - 2304 Cuda Cores - PCIe 3.0 X16 - 1005 MHz - Multi Monitor (Renewed) with tax was $730 on amazon.  

so at that price it is looking like $27,000 for the smallest version of gpt-3

 

edit: just for the GPUs

 

additionally, I don't think openAI will give you the model for personal use, and you won't be able to train a model like it without the HUGE dataset thatt they had. keep in mind you need the hundreds of gigs of vram just to run the model not train it.

 

 

If you want to learn about AI and neural networks you don't need anything other than the computer you have now. you can start training basic networks with a cpu.

 

[igormp]

1 hour ago, AllGoodNamesRGone said:

I started to look at GPU's with CUDA-enabled

That's any Nvidia GPU.

1 hour ago, AllGoodNamesRGone said:

At the end of this month is the last semester of college, studying python, how cool would it be to have figure this out by then.  

By college you mean high school or university? If university, then it worries me a bit lol

 

Otherwise, if you're in high school, start learning about machine learning and make use of google colab to learn, no need to build a new system if you have no idea about what you're doing, it'll be just a waste of money and time.

 

FWIW, GPT-3 is a 800GB model, you can't train/fine-tune it in a regular computer, but will rather need a cluster of those with many V100s/A100s/H100s. Things like GPT-j do exist and you can fine tune those for your need, but without grasping the basics of neural networks you won't go very far.

 

Start off by toying with something like MNIST and go from there.

 

 

For the curious, here are some time/cost estimations for GPT-3:

 

Over 1000 A100s running full tilt for over a month, around $5M in cost.