Live data from Hacker News

HuggingFace Training Cluster as a Service

huggingface.co

1–10 of 47 posts

Re: HuggingFace Training Cluster as a Service

#3
The other day when they announced more funding there was some speculation here about how they would make money, with someone suggesting it's by driving users to cloud gpu platforms (aws, azure). This support that, and it suggest where they will end up, i.e. as a front end for azure.

https://news.ycombinator.com/item?id=37250647

Re: HuggingFace Training Cluster as a Service

#4
At the moment of writing the cost estimate for 70B multimodal model with 7T tokens on 1000 H100 GPUs is $18,461,354 with 184 days of training time.

Anyone willing to share an estimate how cost will come down each year as hardware keeps improving and possible new methodologies are found?

Personally I would not be surprised if it is possible to train the same dataset for half the cost 12 months from now.

Re: HuggingFace Training Cluster as a Service

#5

At the moment of writing the cost estimate for 70B multimodal model with 7T tokens on 1000 H100 GPUs is $18,461,354 with 184 days of training time. Anyone willing to share an estimate how cost will come down each year as hardware keeps improving and possible new methodologies are found? Personally I would not be surprised if it is possible to train the same dataset for half the cost 12 months from now.

Well if you select "trainium nodes", it's already "only" $11,085,287.

Re: HuggingFace Training Cluster as a Service

#8
post #7
post #2

> Train your LLM at scale on our infrastructure Is it really their infrastructure or are they using a cloud provider and this wraps it up and provides convenience for a price?

What’s the difference?

You end up paying more in the latter instance.

Re: HuggingFace Training Cluster as a Service

#10
post #2

> Train your LLM at scale on our infrastructure Is it really their infrastructure or are they using a cloud provider and this wraps it up and provides convenience for a price?

I'd think "infrastructur" includes the nice front end and Python API that they have proven to be capable to pull off already.
Post reply on HN