Tesla's wafer-sized Dojo processor is in production
Tesla's wafer-sized Dojo processor is in production
Tesla-Answer is a dedicated community forum for Tesla owners and enthusiasts. Our platform connects experienced Tesla drivers who share practical knowledge, troubleshooting tips, and maintenance insights to help you maintain and upgrade your Tesla.
Tesla's wafer-sized Dojo processor is in production
I don't really understand the EE side of this or, frankly, the hardware side.
Why is Tesla designing their own chips? How could it be that this makes more sense than buying similar products from companies that specialize in this?
They explained why when the announced they were doing there own chips. From what I remember, it would reduce costs, reduce supply chain risks, keep their AI technology in house, and they need chips designed a certain way to allow for dealing with processing speed bottle necks for their vision system.
Thanks to CHIPS act whomever can mass produce high performance chips in the U.S. is going to make a morbillion dollars
lol because chip design are not that hard for a near trillion dollar company. all the new Tesla cars have their own in house designed chips. Tesla hired a bunch of chip designers from Apple.
Yeah can’t be much harder than getting automatic wipers working.
This infuriates me so much. Like don’t take physical buttons away unless the automated version is nearly fool proof.
Don’t question. Just buy more stock!
18000 amps?! 15kw dissipated power?! And the system is considered "particularly energy efficient"?!
That strains credibility to the breaking point, to be frank. Are there any independent sources?
EDIT: It seems 5x5 D1 processor array "training tile" that actually uses 52v, 288A (15kw of power) refers to previous iteration of Dojo with 7nm tech.
18000 amps seems like a glaring error, so I'd take those numbers from an article with a grain of salt.
Cerebras' wafer-scale processor has peak power rating 23 kW, for what it's worth.
Indeed. But also, Cerebras doesn't seem to be doing terribly well compared to GPUs or TPUs, and Dojo at best is going to be a half-baked Cerebras chip a half-decade behind.
Well, than power draw certainly makes sense (if not 18000 amps, that's insanity). Maybe it is more efficient so far as flops per watts is concerned...
The numbers are actually reasonable. The core voltage is probably 0.8V. 15kW total over 25 processors is 600W/processor. 600W at 0.8V is 750A per processor, or about 18000A for 25 processors.
But is it possible to have amps this high with voltage this low?! That's like a short circuit though a hunk of solid copper bar! I find this extremely hard to believe.
But than, core voltages do get lower and lower and 52v mentioned in an other article is very unlikely to be core voltage, likely PSU voltage, so it gets converted to lower voltage and you do need to multiply V by I to get watts...
If this is true, mere power supply logistics are mind-boggling, let alone cooling solutions.
It's correct, and yes power distribution is a huge challenge for modern chips.
You must be terrified how much CO2 Amazon is omitting in comparison to a one man online seller.
Very cool. I was just hearing speculation about how dojo must be dead
If Dojo is going perfectly and in fact the wafers are shipping now, why is Tesla buying so many GPUs, and why aren't they announcing any performance numbers that matter? This still just sounds like another Itanium. (Itanium shipped lots of chips too. In fact, it was fabbed until 2021!)
Yeah I agree. There’s a middle ground here between things going great and not happening at all though. Given Elon’s track record in AI I have difficulty believing this is going to be groundbreaking or anything, but in the chip domain at least known physics pretty much rules.
Edit: hmm maybe you have more of a point, but I don’t understand why they’d actually manufacture the chip if they know it’s not worth it.
but I don’t understand why they’d actually manufacture the chip if they know it’s not worth it.
Elon has a well-earned history of being 'strategic' about living up to contracts, and anyway, it seems like it would be very foolish to make a large contract like this for an extremely niche chip without payment up front (who would you sell the chips to if Tesla welshed and you seized the chips...?); so I would assume Tesla already paid for these chips a while ago to reserve capacity, and it's use-it-or-lose-it. The payment is a sunk cost, and so the only question is whether you'd like these free Dojo chips that will cost you $0 right now.
Then you'd take delivery because hey, maybe with the hardware, it can work out. Or you just haven't admitted, at the top most levels, that Dojo is a wash and it's time to kill it, and so that means it will keep on trucking (ahem).
And also the PR value. What's the PR value, to Tesla hype, employment/recruiting, and share price, of announcing you are accepting the shipments of Dojo chips? A hard to measure intangible, but it's gotta be at least tens of millions of dollars, and so well worth it kicking the can to another day.
Yeah that all seems plausible. I still put this as a slight Bayesian update towards “maybe it’s working” from my previous state of “it’s almost definitely dead”.
I would agree with that. I still think it's dead, but not quite as strongly as before.
I thought so too