Looks like they are attempting a generic solution like GPT but for vision instead of text.
https://twitter.com/karpathy/status/1365734030247747584
OpenAI GPT-3 and Prompt Engineering
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable
- Thread starter AudubonB
- Start date
Looks like they are attempting a generic solution like GPT but for vision instead of text.
https://twitter.com/karpathy/status/1365734030247747584
OpenAI GPT-3 and Prompt Engineering
Agree in general but I ordered an X on December 7 and there is no existing inventory. I am expecting delivery in April.
asburgers
Sell order in at $8008.5
I feel like we are posting the same tweets multiple times on the same page. Must be getting late.
IMO not widely known, means very few people even in the Tesla community have experience of the FSD beta.
Fewer people still have seen the next version or fully understand how it works.
Posters on TMC have some ideas, but even here opinions are divided, that is because part performance and past track record are not a reliable guide to future performance.
The I guess he is saying solving vision the way Tesla has, and the general pool of NN expertise they have developed, has a wider applicability outside of FSD.
So what Elon knows is significantly more than we know, we know more than than the average Tesla driver, the general public is totally clueless and probably thinks FSD will never work, or is not even aware Tesla is working on it.
In the long run you are right, but I would not put a timetable on it.
The Plaid Model S with the 18650 battery hits some good specs, the whole refresh package, with the yoke etc, is sufficiently innovative and builds on the strengths of the existing design. So there is no rush.
It will take time to ramp 4680 cell production, Berlin and Austin Model Y, the Semi, Roadster and the Cybertruck have to be high priorities
As Yoda would say ..... . "doing the shorts slowly, we are..."
I can’t wait until fsd is released. That is when I think a lot more people will wake up and realize Tesla has won. They will buy the stock and get the car. I think when fsd beta is released to most Tesla owners this will be a huge catalyst for the stock. Also it’s going to be a lot easier having conversations with friends about Tesla. That’s going to be nice cause I’m sick of fighting FUD.
Mo City
Active Member
What about the Semi and Cybertruck? I think it's more likely they get 4680s before non-plaid S & X.
Artful Dodger
"Neko no me"
Samsung considering four sites in U.S. for $17 billion chip plant: documents | reuters (2:02am EST)
It's sillly to believe Tesla needs vertical integration at the Fab level to have a secure supply of FSD chips. Neither bad luck nor poor wx respect the name on the outside of the building.
Cheers!
"Samsung Electronics Co Ltd is considering two sites in Arizona and another site in New York in addition to Austin, Texas, for a new $17 billion chip plant, according to documents filed with Texas state officials."
It's sillly to believe Tesla needs vertical integration at the Fab level to have a secure supply of FSD chips. Neither bad luck nor poor wx respect the name on the outside of the building.
Cheers!
Last edited:
StealthP3D
Well-Known Member
Tesla FUD won't go away, it will just morph
From:
Tesla may talk big but they are only a car company with a showman for a CEO that makes crappy cars that no one wants so they are going bankrupt...
To:
Tesla is out of control, they have monopolized cars, energy production and storage, automated manufacturing, mining and multiple other smaller industries. They have far too much money and power and will take over the solar system if world governments don't band together and stop them.
I personally have a sinking feeling on mediocre Q1 production numbers. They will still be great when you look back, but I'm expecting under 170k and the market to pile on FUD about missing 50% growth.
The only meaningful measure is cumulative YTD compared with previous year YTD. Tesla delivered ~88400 vehicles in 20Q1, 170k would represent 95% yoy growth, to date, which is quite astonishing, no? I would expect that to average down over the year, but given that all factories are ramping this year...
LiveLong&Profit
Member
Some of us long term holders are acting like spoiled children: We want FSD and we want it now.
Because once FSD is out, and people can see it and feel it, and the path towards robotaxis is clear, we know what it will do to make the world better - and the stock prize higher. We like.
Yes, Elon is optimistic and he has a history of estimates which didn't pan out.
None of that matters.
Take a couple of steps back and realize how big a prize FSD is - give the man some slack.
Solving FSD: Big. Huge. FanForkingTastic!
Solving 'a major part of real world AI' (Quote: Elon): Even bigger! Even huger! Even fantastic'er!
Now, if in fact it turns out that the reason why a future FSD is going to work is that it also solves a good chunk of AGI or provides a step change in the path toward AGI : WOW!
Difficult-to-explain-how-huge. Everest huge.
Singularity-huge.
(AGI: Artificial General Intelligence)
Last edited:
Artful Dodger
"Neko no me"
NASDAQ-100 Futures up about 0.87%
Yes.
And it is likely that the chips are made in batches, perhaps 10,000 wafers at a time, so 1-2 million chips or about a years supply. It takes time to set up the steppers with masks for each layer and there are something like 15 different layers required. Initial batches were probably smaller, but that starts getting more expensive. After chip manufacture, there are lots of processes from cutting up the wafers to packaging and testing, this is frequently done in different countries to take advantage of lower wages. Then there is PCB manufacture and another lot of testing.
UkNorthampton
TSLA - 12+ startups in 1
Though I haven't worked in a fab, I did work for a company that runs several fabs. The fabs run 24 hours a day - I assume that is true of any fab that makes chips. My understanding (but I could be wrong) is that these run 24 hours a day for two reasons. The obvious one is that the equipment in the factory is so bloody expensive that letting it sit idle for 8 or 16 hours a day may well be the difference between profitability and bankruptcy. At minimum the equipment is amazingly expensive.
The other reason and more germane here - I think that there are any number of process steps that are effectively continuous flow processes. Continuous flow production isn't happy about stopping.
Clean rooms are also hard to maintain in that state. What I remember reading years back is that the air in the clean room is fully exchanged every minute or two. "dust" in the air gets sucked out of the room fast. My guess is that losing the air flow to the room by itself will cause a delay in starting back up.
Dear lord I surely hope not!
The sorts of complex chips that we're talking about here have either 3 or 4 manufacturers remaining in the world. That is the fruit of decades of consolidation as each new technology node is so much more expensive than the last, that fewer and fewer companies had the necessary volume to make their manufacturing facility economic. So expensive that the volume needed to keep the factory running full time is only available at those few companies in the world. And I fully expect that number to shrink in the next few years as the newer nodes keep getting more expensive (the economics make it inevitable).
Those companies are Intel, Samsung, and Taiwan Semiconductor (TSMC). The 4th that I just don't know enough about is Global Foundries (AMD's former manufacturing facilities). Samsung and TSMC have the volume to be competitive and profitable due to their foundry work - they build other people's chip designs (such as Tesla's chip design). Samsung makes their own designs as well as other companies. TSMC is pure foundry.
Intel is the only remaining vertically integrated design and semiconductor manufacturing shop in the world (which is what Tesla would become as their own chip manufacturer). At the processor level Intel is doing hundreds of millions of units per year - Tesla is going to need a whole lot more demand / volume (I think 2 orders of magnitude, and possible a 3rd) to make their own chip manufacturing economic. Which also assumes that the chips Tesla needs are equally valuable as Intel's chips - some of Tesla's will be worth more, but many won't. That'll further drive the need for more volume. But hey if you've got the volume then there is a lot of value in vertical integration (co-optimize the design and manufacturing technology).
What Tesla can do (and any other design shop in the world) is pony up more money to their manufacturing partner - make sure they have a priority on wafer starts just by spending more for them.
My disclaimer: I used to work for Intel. These are my opinions, beliefs and understandings; don't put any of these on Intel.
Very interesting info. I'll just play out some ideas though
- Mars needs its own chip making
- Many ways to achieve this
- On Earth, relying on free trade, friendly geopolitics & 4 chip makers seems risky
- Elon is adept at solving complex problems
- Tesla/Grohmann etc seem to be innovative, able to make their own machines
- Tesla manufacturing seems to be headed towards extreme space utility & automation
- Musk's companies seem to have most of the jigsaw pieces/skills needed including materials science
- Per gigafactory/country sized chip making not economic now - but could it be?
- FSD, Dojo, Neuralink, Starlink might have enough demand eventually
- A SpaceX/Tesla chipmaker might work with Tenstorrent / Jim Keller or similar - doing low volume AI specialist work that other chip makers might not want initially
- SpaceX/Tesla could decide to bulk out its production with near commodity chips
- SpaceX/Tesla could decide to launch cheap phones or other device to bulk out demand/production
- USA factories might be supported by USA government for defence reasons, paying for a lot of the setup
- Ditto EU
Though I haven't worked in a fab, I did work for a company that runs several fabs. The fabs run 24 hours a day - I assume that is true of any fab that makes chips. My understanding (but I could be wrong) is that these run 24 hours a day for two reasons. The obvious one is that the equipment in the factory is so bloody expensive that letting it sit idle for 8 or 16 hours a day may well be the difference between profitability and bankruptcy. At minimum the equipment is amazingly expensive.
The other reason and more germane here - I think that there are any number of process steps that are effectively continuous flow processes. Continuous flow production isn't happy about stopping.
Clean rooms are also hard to maintain in that state. What I remember reading years back is that the air in the clean room is fully exchanged every minute or two. "dust" in the air gets sucked out of the room fast. My guess is that losing the air flow to the room by itself will cause a delay in starting back up.
Dear lord I surely hope not!
The sorts of complex chips that we're talking about here have either 3 or 4 manufacturers remaining in the world. That is the fruit of decades of consolidation as each new technology node is so much more expensive than the last, that fewer and fewer companies had the necessary volume to make their manufacturing facility economic. So expensive that the volume needed to keep the factory running full time is only available at those few companies in the world. And I fully expect that number to shrink in the next few years as the newer nodes keep getting more expensive (the economics make it inevitable).
Those companies are Intel, Samsung, and Taiwan Semiconductor (TSMC). The 4th that I just don't know enough about is Global Foundries (AMD's former manufacturing facilities). Samsung and TSMC have the volume to be competitive and profitable due to their foundry work - they build other people's chip designs (such as Tesla's chip design). Samsung makes their own designs as well as other companies. TSMC is pure foundry.
Intel is the only remaining vertically integrated design and semiconductor manufacturing shop in the world (which is what Tesla would become as their own chip manufacturer). At the processor level Intel is doing hundreds of millions of units per year - Tesla is going to need a whole lot more demand / volume (I think 2 orders of magnitude, and possible a 3rd) to make their own chip manufacturing economic. Which also assumes that the chips Tesla needs are equally valuable as Intel's chips - some of Tesla's will be worth more, but many won't. That'll further drive the need for more volume. But hey if you've got the volume then there is a lot of value in vertical integration (co-optimize the design and manufacturing technology).
What Tesla can do (and any other design shop in the world) is pony up more money to their manufacturing partner - make sure they have a priority on wafer starts just by spending more for them.
My disclaimer: I used to work for Intel. These are my opinions, beliefs and understandings; don't put any of these on Intel.
Thanks. In addition to the well-known clean rooms, these fabs also have significant water consumption. And with their product having to have a very high purity, the water will surely go through the plant's own processing.
Still, the information related to Samsung's Texas plant does not add up:
1) Fremont was actually closed - but only for two days before production was running again, according to both Elon Musk and on-site social media reporting.
2) The Texas Samsung plant was reportedly shut down in an orderly manner. This would presumably include that any water processing facility was also properly shut down (thanks @truth_tesla),
https://twitter.com/truth_tesla/status/1364926797708750851
3) So the Korean news that the Texas Samsung plant should not only have been shut down until now but continue to be so until mid-April (due to water problems), does not make sense.
PS. I imagine Samsung being intensely negotiating deliveries with their customers - surely Tesla should be able to give an interesting offer of some battery-powered back-up. Maybe not for the mentioned 150MW consumption for very long, but at least enough to ensure that proper shutdown procedures can be followed (or maybe Samsung has a fossil-powered generator on-site that needs a few minutes to spin up).
Last edited:
Dave Lee and James Douma discuss how what Tesla is doing with their birds-eye-view (BEV) neural nets is novel and will likely have other non-FSD applicability.
Note: Dave has two recent videos with James discussing the topic. Both are valuable to understand what Tesla is doing in this space.
Shutdown likely went fine. It is the startup that likely went off the rails. It is a guess but my guess is that the fly in the ointment is the chemical cocktail that came down the pipes after the startup.
Too much chlorine (beyond design specs) and a very small molecule can overwhelm the activated carbon prefilters ahead of the chemically sensitive RO membranes and then the RO is damaged and “stuff” gets into the pipes. Now there is contamination in the system and hard to clear. My understanding is that Fabs require ultra pure water.
This may all be wrong because I don’t know the Samsung’s Fabs specifics at all. Maybe they have their own wells for instance. But an infrastructure failure to the water system like parts of Texas experienced can result in all sorts of “stuff” back flowing into the pipes.
Many states are enacting water regulations requiring back flow preventers on all commercial facilities to minimize unknown “stuff” from back flowing into the mains during low pressure events. Could Texas be slow on adopting such regs?
In any case Samsung will likely quickly clarify things. Maybe Ron Baron will know something on Thursday.
So you know what actually happened?
They are reporting that they can't restart because of "water problems". What are those problems? Did equipment freeze with water in it and get destroyed? Did contaminated water get in the equipment when the water treatment plants were down and damage the equipment? Did pipes freeze, burst, and flood parts of the factory?
I trigged on the water aspect. Full disclosure: I am a little invested in a Swedish startup conglomerate making water purification plants for various use cases and scales. One of them is Xzero (and the new spinoff Type1) which produce ultra-clean water specifically for microchip making and the like. Available in sample quantities now, I think.
So, if rinse water is the narrow sector for Samsung Austin, maybe they should try out a few bottles and/or put in an order for a hitech still on-site? Similar applies to others on a lab scale.
EDIT: I now see that @lascavarian threw his hat in too.
The Samsung plant is a billion dollar facility. While mistakes do happen, they probably have pretty good procedures for how to safely restart production - created from decades of experience. And unlike the shutdown, the restart should not be under time-pressure due to the risk of damaging the facility by being too slow.
PS. Also, reports are that production is actually running, so these reports are hard to reconcile.
Similar threads
- Locked
- Locked
- Locked
