Tesla Developed Its Own GPU Transfer Language

in #hive-16792218 days ago

▶️ Watch on 3Speak


When it comes to AI companies, Tesla is right that. Many wil doubt this and still call it a automotive company.

In this video I discuss how Tesla creates a transfer protocol for its GPUs that is twice as fast as the industry standard, which was developed by Nvidia. This means Dojo, running on Tesla's own chips will have an advantage over the speed transfers for its Nvidia cluster.


▶️ 3Speak

Sort:  

Tesla dont want to be dependent on Nvidia for their Gpu needs.

They are still buying and have an order in for Blackwell processors.

The problem is Nvidia isnt going to deliver enough to meet their needs.

That's quite amazing and it is nice to see from the sidelines how companies are transforming themselves by being where the money and technology is. Being able to pivot in so many directions is imperative in order to thrive nowadays.

Those that transition are going to be in the game down the road.

I have a feeling many will not do it and end up being disrupted.

I don't think I've seen much coverage of this. How much edge does this give Tesla over their competitors? I know they still use similar GPUs so the compute is similar, but how often does this transfer protocol come into play in the training? If it is frequent, and it is twice as fast as the industry standard, does that mean Tesla will be getting a sizeable lead soon?

What competitors?

Well I am not sure there is competition in that field at this point. Everyone is trying to stake their markets. It is so early there are no established players other than the hardware entities.