BYD are possibly in second place even if they need a telescope to see Tesla, I can't think of another strong candidate for second place.
First, second and third all get a medal, we don't want Tesla standing on the podium by themselves.
-
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
First, second and third all get a medal, we don't want Tesla standing on the podium by themselves.
insaneoctane
Well-Known Member
Probably as soon as their owners start plugging them in.
Words of HABIT
Active Member
Courtesy of The Motley Fool. If history repeats itself the lows will be in September:
Statistically, these are Wall Street's worst-performing months
But there's another side to this story. While July, April, December, and January have historically been kind to investors, another three months have, statistically, been a net negative for Wall Street.Since 1928, the following are the only three months to produce an average annual loss for the S&P 500:
- September: (1.1%)
- February: (0.1%)
- May: (0.1%)
IMO Tesla would be reluctant to take a path that means FSD doesn't work on HW3 and can never work on HW3.
That doesn't necessarily rule out using the second NN processor while FSD is supervised, and a suitably qualified human can take over.
For fully redundant unsupervised FSD, they need to option of failover to the second processor.
So for Robotaxis using HW3 it eventually needs to fit. (Or in the event of a problem the Tesla needs to safely park itself.)
I've always been slightly sceptical about claims that the second HW3 NN processor is being used because the NN can't fit on the first processor. I think it was being used, but we don't really know why, it could simply be gathering targeted fleet data.
These are areas where so far Tesla makes no official comment, so what we have is informed speculation which is not always 100% right.
Last edited:
Words of HABIT
Active Member
Stephenson Line held for 2 1/4 years
Hoping the parallel WoH Line (taking into account recent bottom supports) can hold for at least that long. If so, TSLA must remain above $210 and and growing daily.
Either way, trend is clearly upwards.
Hoping the parallel WoH Line (taking into account recent bottom supports) can hold for at least that long. If so, TSLA must remain above $210 and and growing daily.
Either way, trend is clearly upwards.
It still makes a lot of sense because people still have to plug them in. How long and how many fines did it take for most people to plug in their seatbelts? The main reason people purchase a PHEV is HOV lanes. Most don't care about the battery part.
Artful Dodger
"Neko no me"
Chinese owner of iconic MG car brand to build Europe plant - BBC www.bbc.com › news › technology-66117178 | Jul 06, 2023
"China's largest car manufacturer SAIC Motor says it will build its first factory in Europe, after sales of its vehicles on the continent jumped. The state-controlled company - which owns the iconic MG brand - says the new plant will produce electric vehicles."
These will sell well in Europe if they are eligible for Eurozone incentives:
Cheers!
As we all know here….from the point of saying your going to build a factory to the time which cars are rolling off of the assembly line is a long….long time period.
Tesla tried to move as fast as possible and they were promising such a large factory that you’d think they would have gotten things fast tracked and it still took 2 years
dhanson865
Well-Known Member
Is it truly parallel or just close to parallel?
I still expect us to get above Stephenson again someday so if they are truly parallel we'd pass both.
Uber burned cash from inception, so did Amazon. It's good to make profits but not really a requirement. Dont believe me? Look at corporate Japan which rarely if ever produced profits in key industries in the 80s and into 90s . They invested in market share and really that is still the goal from what I glean from my relatives. What is key is having enough cash to make it to some point where your losses stop. Uber maybe just got there.
Waymo will not run out of money, google can fund Waymo with 2 qrt profit until the point I would predict cash flow neutral. Now that Cruise has some velocity I would be astounded if they run out of funds. I am sure Tesla gets to working FSD just not sure when.
I view Waymo as more dangerous, Google has lots of reasons to want captive non driving consumers in a vehicle. Ad money for restaurants, for bars, for groceries, etc. Nobody makes more from ads than google and this gives them a new market.
Oh, so you were the guy in the Bay Area with the horrendous “REX ENVY” license plate.
First place is BMW, 2nd is Mercedes and Tesla is third, tho about to dethrone Mercedes. BYD is not in the luxury segment and it's ranked near the bottom in the mainstream segment no one even cares about them.
People spend too much time segregating ev vs ice when it's all just metal boxes that transport people.
Knightshade
Well-Known Member
It's not. For one thing, we've known for years only one of the two nodes can do campaign triggers at all (gathering fleet data)- they're not entirely identical you see. For another- Both Green and James Douma have confirmed it's actually them needing to cross nodes because the NN stuff can't fit on one node- not for years now.
The "cost" in terms of compute by crossing dies is high, there'd be no reason at all to eat that cost if they had a choice.. (they'd instead run the "active" NNs on one die, and any "gathering data" code on the second die). Instead they have code using extended compute that crosses dies, and have ever since they ran out of fits-in-one-node stuff several years ago on HW3)
This isn't speculation- here's green remarking on it and also mentioning what Douma found:
again that's from over 2 years ago- the code hasn't gotten LESS intensive since, it's gotten moreso, with significantly more extended-compute request stuff (ie things it can not run in one node and is forced to take the performance hit of crossing nodes to borrow from the other nodes NPU.)
Ultimate- nobody knows "how much is enough" for generalized L5 driving until someone actually achieves it-- Tesla has been wrong about the answer at least 3 times now (HW2, HW2.5, and HW3- all of which they claimed were enough and clearly are not)-HW4 may be enough- it also may NOT be-- because anyone who says they know how much IS is just guessing until they actually do it.
FYI- one of the "big" deals in Dojos design was the super fast interconnects so you do NOT take the normal "this has to leave the die" penalty most stuff does- but that's no help to HW3 (or HW4 for that matter though each node has more compute than the total HW3 system had- but should they need to cross nodes there eventually they'll see the same types of cross-die performance penalties because the basic architecture hasn't changed).
Again all this has been covering, in detail, for years now, in various threads over here:
AI, Autopilot, & Autonomous/FSD
Discussion about AI, Tesla Bot, Tesla Autopilot (AP), the promise of Full Self Driving (FSD), as well as other Autonomous Vehicles.
teslamotorsclub.com
Anybody factoring FSD into their investment thesis and not current on this stuff and what's in that forum are doing themselves a great disservice simply speculating in here on stuff we already know pretty well.... (doubly so for anyone who isn't and hasn't been actively testing FSDb for a while)
Last edited:
There isn't a lot of evidence that BMW, and Mercedes can translate their success in the ICE market to the EV market,
There is also not a lot of evidence that the ICE market will be relevant in 5-7 years time.
I'm increasingly pessimistic about legacy auto, I expected them to be doing better, and in many ways it is disappointing how badly they are doing with the move to EVs.
Buckminster
Well-Known Member
Not saying FSD computer didn't run out of compute 2 years ago, but this back and forth doesn't really matter as they optimize how they utilize their resources and have IMPROVED FSDB drastically over the last 2 years. "The code is now more compute intensive today" implies that FSDb would be underperforming as it's being choked by the hardware that was obsolete 2 years ago but this is simply not the case. Performance drastically decrease if there's a severe bottleneck in the hardware as it runs out of resources. When a game runs out of Vram by 10%, frame rate drops to single digits...it doesn't just slow down by 10%, but by 95%. So if current code requires 120% of what the FSD computer can give, then FSDB wouldn't even work. It doesn't work @80% until there's enough compute in HW4 to make up the difference.It's not. For one thing, we've known for years only one of the two nodes can do campaign triggers at all (gathering fleet data)- they're not entirely identical you see. For another- Both Green and James Douma have confirmed it's actually them needing to cross nodes because the NN stuff can't fit on one node- not for years now.
The "cost" in terms of compute by crossing dies is high, there'd be no reason at all to eat that cost if they had a choice.. (they'd instead run the "active" NNs on one die, and any "gathering data" code on the second die). Instead they have code using extended compute that crosses dies, and have ever since they ran out of fits-in-one-node stuff several years ago on HW3)
This isn't speculation- here's green remarking on it and also mentioning what Douma found:
again that's from over 2 years ago- the code hasn't gotten LESS intensive since, it's gotten moreso, with significantly more extended-compute request stuff (ie things it can not run in one node and is forced to take the performance hit of crossing nodes to borrow from the other nodes NPU.)
Ultimate- nobody knows "how much is enough" for generalized L5 driving until someone actually achieves it-- Tesla has been wrong about the answer at least 3 times now (HW2, HW2.5, and HW3- all of which they claimed were enough and clearly are not)-HW4 may be enough- it also may NOT be-- because anyone who says they know how much IS is just guessing until they actually do it.
FYI- one of the "big" deals in Dojos design was the super fast interconnects so you do NOT take the normal "this has to leave the die" penalty most stuff does- but that's no help to HW3 (or HW4 for that matter though each node has more compute than the total HW3 system had- but should they need to cross nodes there eventually they'll see the same types of cross-die performance penalties because the basic architecture hasn't changed).
Again all this has been covering, in detail, for years now, in various threads over here:
AI, Autopilot, & Autonomous/FSD
Discussion about AI, Tesla Bot, Tesla Autopilot (AP), the promise of Full Self Driving (FSD), as well as other Autonomous Vehicles.
Anybody factoring FSD into their investment thesis and not current on this stuff and what's in that forum are doing themselves a great disservice simply speculating in here on stuff we already know pretty well.... (doubly so for anyone who isn't and hasn't been actively testing FSDb for a while)
Many people speculated that HW3 FSD computer was no where near powerful enough to achieve today's FSDb performance 2 years ago because they speculated Tesla was already on the edge of computational power and FSDb performed like ass back then. "If FSDb couldn't drive in a straight line today at 90% HW3 utilization, how will it ever drive straight without HW5 or 6?".
Last edited:
Questione: in PHEVs, is the regen brake efficient enough to supersede the fact that you carry around a battery/mass? Many people don't plug in but they recharge the battery with the brake alone.
George Hotz manages to run his 3 camera input end-to-end NN on a mobile phone CPU (Qualcomm Snapdragon 845 for crying out loud) and drive to the nearest Taco Bell. So I highly doubt these "experts" who say that Tesla FSDBeta cannot run on one node.It's not. For one thing, we've known for years only one of the two nodes can do campaign triggers at all (gathering fleet data)- they're not entirely identical you see. For another- Both Green and James Douma have confirmed it's actually them needing to cross nodes because the NN stuff can't fit on one node- not for years now.
The "cost" in terms of compute by crossing dies is high, there'd be no reason at all to eat that cost if they had a choice.. (they'd instead run the "active" NNs on one die, and any "gathering data" code on the second die). Instead they have code using extended compute that crosses dies, and have ever since they ran out of fits-in-one-node stuff several years ago on HW3)
This isn't speculation- here's green remarking on it and also mentioning what Douma found:
again that's from over 2 years ago- the code hasn't gotten LESS intensive since, it's gotten moreso, with significantly more extended-compute request stuff (ie things it can not run in one node and is forced to take the performance hit of crossing nodes to borrow from the other nodes NPU.)
Ultimate- nobody knows "how much is enough" for generalized L5 driving until someone actually achieves it-- Tesla has been wrong about the answer at least 3 times now (HW2, HW2.5, and HW3- all of which they claimed were enough and clearly are not)-HW4 may be enough- it also may NOT be-- because anyone who says they know how much IS is just guessing until they actually do it.
FYI- one of the "big" deals in Dojos design was the super fast interconnects so you do NOT take the normal "this has to leave the die" penalty most stuff does- but that's no help to HW3 (or HW4 for that matter though each node has more compute than the total HW3 system had- but should they need to cross nodes there eventually they'll see the same types of cross-die performance penalties because the basic architecture hasn't changed).
Again all this has been covering, in detail, for years now, in various threads over here:
AI, Autopilot, & Autonomous/FSD
Discussion about AI, Tesla Bot, Tesla Autopilot (AP), the promise of Full Self Driving (FSD), as well as other Autonomous Vehicles.
Anybody factoring FSD into their investment thesis and not current on this stuff and what's in that forum are doing themselves a great disservice simply speculating in here on stuff we already know pretty well.... (doubly so for anyone who isn't and hasn't been actively testing FSDb for a while)
IMO, those 3 answers are basically the same, maybe just focussing on different aspects. I’ll try to rephrase my answer:
Image that 300K lines of C++ code as navigating through a giant maze, carefully deciding at each step what step to take next. It runs on the CPU part of the autopilot computer, because that part is really good at such code (i.e. evaluating something and then deciding which step to take next). It is developed by carefully thinking of all the possible situations that might arise, and writing the code corresponding to those situations and what to do next. The output is e.g. the speed at which to drive, which angle the steering wheel should be at, etc.
Imagine the neural network part as a firehose of calculations, doing many calculations at once. It runs on the part of the autopilot computer that is optimised to do hundreds or thousands of calculations at once. It just adds and multiplies numbers until it outputs the same speed to drive and angle to turn the wheel. It is developed by training a neural network by providing thousands of samples of situations and what the speed and angle should be.
Replacing the C++ code with neural network will lessen the load on the CPU part of the autopilot computer, and increase the load on neural network part of the computer. Only Tesla knows how this affects the available neural network resources on the autopilot computer: While they are adding load on the neural network part of the computer because of the added V12 functionality, they may also optimise (i.e. lessen) the load used by the V11 functionality.
petit_bateau
Active Member
Thank you @NicoV . That is helpful. So bottom line is we don't yet know if eliminating the 300k lines C++ will enable it to fit on one processor in HW3. With either the non-optimised or (future) optimised NN. As always FSD is still a carrot out of reach.
Similar threads
- Locked
- Locked
- Locked
