Will HW2.5 computers taken from upgraded cars be put to use in Tesla's "Dojo" supercomputer?

[ratsbew]

I have a hard time believing that Tesla is just going to throw away tens of thousands of very high performance Nvidia GPUs as they take in HW2.5 units from cars that are being upgraded to HW3.0.

My money is on them taking all of these HW2.5 units and using them as their supercomputer that is to be known as "Dojo". If not Dojo, I am almost certain that they will use them for some kind of supercomputing whether it is internal or rented out. All of those HW2.5 units combined would be one of the fastest computers on the planet.

 

[diplomat33]

I have a hard time believing that Tesla is just going to throw away tens of thousands of very high performance Nvidia GPUs as they take in HW2.5 units from cars that are being upgraded to HW3.0.

My money is on them taking all of these HW2.5 units and using them as their supercomputer that is to be known as "Dojo". If not Dojo, I am almost certain that they will use them for some kind of supercomputing whether it is internal or rented out. All of those HW2.5 units combined would be one of the fastest computers on the planet.

I am picturing some mega supercomputer that becomes sentient. HA HA. Maybe that is why Elon is always saying we should be afraid of AI.

 

[heltok]

Very different systems for inference and for training. For inference you need much less memory and bandwidth than for training. Also very different to design a cloud server and to design an embedded system.

Would be very surprised if Tesla are using them for anything else than spare parts or to recycle them.

 

[banned-66611]

Their "supercomputer" isn't a supercomputer. It's a rented cloud service using commercial GPUs with much, much higher performance per note that HW2.5.

Anyway they will probably abandon it soon when they realize that image recognition AI can be done much cheaper and more reliably using an off-the-shelf commercial solution like the one Google has, and when it becomes apparent that just recognizing things isn't going to make FSD work.

 

[clydeiii]

Seems very unlikely.

 

[DrDabbles]

My money is on them taking all of these HW2.5 units and using them as their supercomputer that is to be known as "Dojo"

Don't throw your money away just yet.

The connectivity between all of these compute nodes would be 1 gig ethernet. The memory footprint is very low. The main system CPU is a very low powered ARM unit that would be overwhelmed almost immediately with scheduling tasks and sending data to the GPUs. Simply put, there isn't enough throughput on the network, there isn't enough throughput on the local bus, there isn't enough RAM to effectively cache tasks, and the anemic GPCPU is a huge roadblock. In other words, no. These things won't become a "supercomputer".

Dojo will be run on cloud instances, very likely on existing GPU platforms or with hosted ASIC/FPGA solutions. Sinking the capital required into building your own datacenter would be an unbelievable drag on Tesla's financials, and they'd need to keep updating the hardware every few years to get financial and performance benefits from it. It would be a complete waste, and it's not at all their core competency. Every dollar spent on building datacenter space could be better spent building gigafactories and battery production lines, which is something that actually makes them money today.

 

[banned-66611]

The good news is that there should be lots of cheap HW2.5 units showing up on eBay eventually, providing a handy hobbyist platform.

 

[ratsbew]

I think we found our answer. Tesla is putting used HW2.5 units in Chinese Model 3s. Makes perfect sense and probably shaves $500 off of the production cost by using "flight proven" hardware.

Tesla delivers Model 3 without self-driving computer, blames supply chain issue - Electrek

 

[Mardak]

It seems unlikely that Tesla still has a backlog of HW2.5 inventory nor ordering new HW2.5 from Nvidia, and given that HW2.5 is optimized for inference and not training, it is indeed quite practical to reuse "flight proven" hardware for vehicles that are less likely to need FSD computer. I suppose there's some actuarial science in the trade-off of a new vehicle with HW2.5 installed then needing a FSD computer retrofit.

Although with HW2.0 retrofit to FSD computer, it seems like definitely that older hardware won't make it in to new vehicles, so what will happen to those…

 

[DrDabbles]

It seems unlikely that Tesla still has a backlog of HW2.5 inventory nor ordering new HW2.5 from Nvidia, and given that HW2.5 is optimized for inference and not training, it is indeed quite practical to reuse "flight proven" hardware for vehicles that are less likely to need FSD computer.

Probably worth reading this thread about it. Elon Musk on Twitter

Using hardware you've already got makes the most financial sense.