News Tesla scraps custom Dojo wafer-level processor initiative, dismantles team — Musk to lean on Nvidia and AMD more

[Admin]

Tesla is shutting down its Dojo supercomputer program, reassigning staff and leaning more on AMD and Nvidia. Tesla's eventual designs expected to rely on a unified architecture spanning from edge devices to data centers.

Tesla scraps custom Dojo wafer-level processor initiative, dismantles team — Musk to lean on Nvidia and AMD more : Read more

 

[autobahn]

DOJO wafer got DOGE'ed! I wonder what this means for a company like Cerebras that is working on this same tech. Is there something flawed with the approach of one giant wafer?

 

[Notton]

Is there something flawed with the approach of one giant wafer?

Yeah, pretty much everything.
It's using the entire wafer, so there's no yield rate. Instead it's failure rate, or how much of the silicon can be used.
It's not just the print, but how well it's bonded to a substrate.
How to deliver power, how to keep it running cool, how to prevent damage from warping when heated.
I'm sure there's more, but bigger chips are generally are troublesome compared to small chips. It's why AMD went with chiplets that worked fine with CPUs/APUs, but failed at GPUs.

 

[Bikki]

Elon reasoning is like this:
- Dojo cant match Nvidia hardware and software any time soon
- The cost of developing this is through the roof, Tesla is not big enough for this
- The consumer facing Tesla dont need to be server chip giant
- Inference chip is good cause no-body produce better edge vision inference chip atm, will be sole focus for now

 

[Bikki]

Yeah, pretty much everything.
It's using the entire wafer, so there's no yield rate. Instead it's failure rate, or how much of the silicon can be used.
It's not just the print, but how well it's bonded to a substrate.
How to deliver power, how to keep it running cool, how to prevent damage from warping when heated.
I'm sure there's more, but bigger chips are generally are troublesome compared to small chips. It's why AMD went with chiplets that worked fine with CPUs/APUs, but failed at GPUs.

Very insightful, thank you. I still think the main reason this get the axe is development / usability. (Looking at Cerebras, they’re still going strong). You have very strong alternative hardware for training (Nvidia) and low usage (only 100k h100 deployed for Tesla vs 500k xAI), compare to sky high developing cost just not worth it.

 

[abufrejoval]

I've been expecting this for more than two years now, and for two main reasons:

1. Scale: while DOJO's variable precision design seemed to make it much more generic than many more popular ones, the whole thing really was only made for a single insider job for Tesla. And no single job can sustain the cost of current chip designs, you not only need to sell outside, you need many product shoulders to lean on. I wonder how the latter will turn out for Cerebras, but I can't see them last for many generations, either.
2. Musk's autistic obsession against sensor fusion and reliance on vision-only: yes, humans don't have LIDAR or Radar, but we still do sensor fusion with sound, inner-ear velocimeters, our skin and even our stomachs. And of course we carry a vastly richer model of the physical world within us than can be recorded through cameras alone. Babies may have a richer understanding of physics than DOJO ever did.

Too bad for the environment those chips will go into a landfill without delivering much use, but a least the science seems to have found another place to grow.

 

[fiyz]

Isn't it great being a CEO, and spending other people's money on failed ventures?

 

[pointa2b]

Isn't it great being a CEO, and spending other people's money on failed ventures?

He's out there doing things and taking risks, and more often than not they are successful (hence his net worth).