It used to be 2x number of parameters, so about 350GB.

Facebook came with an optimization a couple months ago so it takes only half the memory, so about 175GB for GPT-3, not sure if OpenAI has implemented that yet.

BLOOM, which is around the size of GPT-3, though not as good currently, works with these optimizations and should then be possible to run with 8x 24GB consumer GPUs.

Huggin face has a nice blog post on this

https://huggingface.co/blog/hf-bitsandbytes-integration

As far as I know Davinci is supposedly a variant of gpt3-175B, so likely 175B parameters, so likely in the hundreds of gigabytes of ram unfortunately

It is not available to run locally, but like the other comments said, an equivalent model would require an inordinate amount of VRAM.

I'm using the text-davinci-003 model with the open AI API. It seems equivalent to chatGPT in terms of coding tasks. And the pricing is totally reasonable.