You can install our site as a web app on your iOS device by utilizing the Add to Home Screen feature in Safari. Please see this thread for more details on this.
Note: This feature may not be available in some browsers.
Not sure how they compete with Nvidia and Dojo without the rest of the system.
Elon described Dojo as still being a long shot.
Plans for v1.5 & v2 etc.
Doing training jobs now
alternate translation: nVidia is a very good competitor and has been able to meet/beat every critical performance milestone for the Dojo project so far (the net: They push each other to go even faster)
One lesson Tesla finally learned, "don't show the secret sauce ahead of time". So this could also be applicable to DOJO stmt.
Things now will all be shared in the almost alpha/production version, not in proto and beta versions.
If anyone doubts that nVidia is competing in a league of its own for AI compute, I suggest you check out the latest training results: Leading MLPerf Inference v3.1 Results with NVIDIA GH200 Grace Hopper Superchip Debut | NVIDIA Technical Blog
Summary: Google's TPUs got knocked down and there are simply no other competitors currently. This is why Dojo isn't being used for production training at Tesla.
Huh? Elon specifically said that Dojo is being used for training.
Exactly, if Dojo costs 25% less than unobtainium Nvidia and is half the speed they still have a winner.They mentioned working on future versions of Dojo... if it was going really badly they would just abandon the project....
Dojo isn't just hardware, there is supporting software, a lot to debug and optimise...
It is a bit like batteries,,,, 4680s didn't mean Tesla would stop using competitors batteries, sometimes the competing product is the best option and the fastest path to the desired result.
Would they start the Dojo project knowing what they know now? That is the difficult question to answer... one risk is that competing product becomes high priced or hard to get.
What is the difference between a bad idea, and idea that takes a lot of hard work over a long timeframe to achieve...?Dojo is one of Elon's few bad ideas, just a step below the Boring Company.
Having an inhouse supercomputer is a great idea. Having it based on your own proprietary chip isn't. Elon gambled and failed.What is the difference between a bad idea, and idea that takes a lot of hard work over a long timeframe to achieve...?
The difference is most people give up far too early, and far too easily, that is why they never achieve anything..
It takes tenacity to make things work...
Elon does a few things that are hard to achieve, or which might never be achieved, which if they can be done have a large pay-off. Someone has to tackle the difficult problems.
I think the project is proceeding on the basis that it may work out, and may come in handy, that was always the plan..Having an inhouse supercomputer is a great idea. Having it based on your own proprietary chip isn't. Elon gambled and failed.
Solar pretty much is all about scale and making them cheap. The solar roof has innovation but still requires work to make the install easier.
When it comes to AI chip it's another story. Amd and Nvidia holds the majority of the GPU ip while TSMC is the world's most advanced fab. The barrier to entry is near impossible as we see chip giant Intel struggle even though they have been making igpu for 2 decades.
However Dojo main ability is its from scratch design with massive focus on interconnects as Tesla predicts gigantic models eventually bottle necking the design from other manufactures. Even though amd has multichip design using infinity fabric and Nvidia using infiniband, both are extremely fast, but cabinets to cabinet interconnect still needs work.
But yea the problem here is cost. 40k for a Nvidia gpu and 15k for an amd one is literally the cost of a car. If Tesla can get one working at half cost then it's a huge win.