MrBadger
Badger out
EM: December for sure, or definitely the month after that.
: But Elon, there is no month after December.
EM: Exactly
: But Elon, there is no month after December.
EM: Exactly
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD V9 updates?
- Thread starter mazers
- Start date
Hey! At least it's now the "next week", not "two weeks"! That's progress. In a year or so we'll be down to "tomorrow".
Knightshade
Well-Known Member
Green suggests he's seeing HW3 is already out of spare compute, forcing Tesla to abandon the redundancy idea and try and use both nodes
So....HW4 anybody? Should be ready soon right?
mikes_fsd
Banned
This is why I laugh at most greens posts where he is not posting direct (factual) data from the firmware.
Remember his rant about how Tesla is replacing the current radar with the 4D radar, just because he saw a new vendor in the firmware???
He loves to jump to conclusions....
But then again, that wouldn't make for a sensational post on Twitter would it?
Last edited:
mikes_fsd
Banned
Tesla purchased DeepScale in 2019, their whole selling point was NN optimization to run on less powerful computers.
I am willing to bet that none of the DeepScale optimization has been performed on the current NN's as they are still trying to get to a certain level of functionality before focusing energy on that effort.
Knightshade
Well-Known Member
This is why I laugh at most greens posts where he is not posting direct (factual) data from the firmware.
Remember his rant about how Tesla is replacing the current radar with the 4D radar, just because he saw a new vendor in the firmware???
But... the vendor being there was him posting direct (factual) data from the firmware.
So I find your objection confusing.
Does SpaceX do a lot of visual perception NN code on multi-node architecture not originally written or conceived to work multi-node in the first place?
Because if not that expertise might not be quite as applicable as you suspect.
He's been citing direct (factual) data from production firmware about Tesla hitting the limits of single node compute since last year, and their struggles to extend compute using the other node, and how poorly its been going in general.
Those are factual.
The only speculation in his entire thread is he's speculating this is the primary reason they keep delaying a new version of the beta.
So they bought a company 2 years ago and haven't bothered, even during what they claim was a near total re-write of code to 4D, and then another one to vision-only, to actually apply the thing they bought them for to that code?
That seems.... not an ideal use of resources.
And exactly the sort of wild speculation you're accusing the other guy of engaging in.
More likely- They had no idea the needed in-car compute when they developed HW3, and they got it wrong.
This is more likely because it's exactly what they already have done with the HW2/HW2.5 setups.
Both of which, like HW3, they claimed was "all" the hardware you'd need for FSD... until they realized it wasn't.
A fair number of folks speculated that'd be the case back when HW4 was mentioned during autonomy day--- that it was in development in case it turned out HW3 wasn't sufficient.
Green makes a pretty good case, again based on the actual behavior of current firmware on HW3, that that will be the true.
The elephant in the glovebox of course is- did they design HW4 to physically swap in place of HW3 as they did before- so when they have to do free upgrades for all FSD buyers it's easy?
mikes_fsd
Banned
I don't see it that way.
It was reported as an acqi-hire to get all the AI engineers from the company, so I can see them using the engineers for the actual FSD development before focusing everyone on optimization.
Even in the smartphone world after so many releases Android had a "project butter" to optimize the responsiveness of the entire OS.
You focus on getting the features built out, then you go over it again to optimize.
Knightshade
Well-Known Member
That's certainly possible.
But it's exactly the sort of speculation you seemed upset at Green making, despite him have actual data from the compute usage in current FW to go on.
The problem is nobody, NOBODY, knows how much on-car compute you need for real L5 FSD. Because nobody has it yet.
It's theoretically possible it can be done on HW3 with sufficient optimization or maybe some other future rewrite or something.
But from the evidence we've got so far, it suggests it won't be and they'll need more compute in the car to get there.
Now my own remarks that veer into speculation-
It seems to me like if they were already hitting walls with the original design intent (which was dual, redundant nodes) that would've been a time to look at optimizing--- instead they seem to have chosen to give up on redundancy and just try and make it work using both nodes running as one instead.
Hopefully I don't need to cover why that's an insanely bad approach for something you honestly expect to ever be an L5 robotaxi solution.
So I think at best the idea there would be "Let's see if we can get something working with ALL the compute of both nodes so at least we'll know how much compute a single node needs in the actual real L5 hardware we will need to develop and upgrade everyone to down the line"
Course, the answer might end up being "more than both nodes together on HW3 can give us" and then they're really in trouble....especially since Keller has at least 2-3 more years before he'd be interested in job hopping back to Tesla.
But it's exactly the sort of speculation you seemed upset at Green making, despite him have actual data from the compute usage in current FW to go on.
The problem is nobody, NOBODY, knows how much on-car compute you need for real L5 FSD. Because nobody has it yet.
It's theoretically possible it can be done on HW3 with sufficient optimization or maybe some other future rewrite or something.
But from the evidence we've got so far, it suggests it won't be and they'll need more compute in the car to get there.
Now my own remarks that veer into speculation-
It seems to me like if they were already hitting walls with the original design intent (which was dual, redundant nodes) that would've been a time to look at optimizing--- instead they seem to have chosen to give up on redundancy and just try and make it work using both nodes running as one instead.
Hopefully I don't need to cover why that's an insanely bad approach for something you honestly expect to ever be an L5 robotaxi solution.
So I think at best the idea there would be "Let's see if we can get something working with ALL the compute of both nodes so at least we'll know how much compute a single node needs in the actual real L5 hardware we will need to develop and upgrade everyone to down the line"
Course, the answer might end up being "more than both nodes together on HW3 can give us" and then they're really in trouble....especially since Keller has at least 2-3 more years before he'd be interested in job hopping back to Tesla.
mikes_fsd
Banned
This is literally been my point as the limiting factor for YEARS. Computer vision can eventually figure out to do a lot, but it may take a massive amount of compute power. I have no idea if HW3, or HW4, will have enough compute and memory to handle the size of the models needed for a certain accuracy.
I feel like there has got to be some room for optimizing the code. First revision of code has to have some fat on it right?
Knightshade
Well-Known Member
Some? sure.
But as I said, I think the fact they didn't bother going back to do that to fit it in a single compute node and instead just abandoned redundant nodes to try and do parallel compute with both tells us they don't think there's enough room, even optimized, to run single node.
If they did then failing to optimize now means they waste a ton of time making it work across 2 nodes needlessly.
Much more likely seems to be they've concluded "We have no idea how much compute we will end up needing- but it's more than 1 HW3 node will give us no matter how optimized the code" and they're hoping they can get a working solution before they max out BOTH nodes now.
If so then HW4 (or HW5 if the need exceeds 4, since 2 years on HW4 has to be pretty far baked already) will have a known target a single node must be capable of.... (since they will HAVE to have everything running in 1 node for redundancy if they want L4/L5).
If they max HW3 even using node B for extra compute and still don't have vision solved they have even bigger problems, because they then have to move everyone onto HW4 for further progress and STILL can't be sure it's sufficient.
Zhelko Dimic
Careful bull
I agree with most everything you said.
Except current buyers of FSD (as of 2019) aren't due redundancy, and I'm not sure if earlier buyers of FSD are due it for free either. If I recall promises correctly. Description of FSD was based on capabilities, not the formal definition of the L4/L5. However, idea that car can find you wherever you are, and that you could sleep behind the wheel may make them have to offer free redundancy for pre-2019 FSD. Was idea of sleeping behind the wheel really a promise, or Elon improvising?
In any case, I feel this is similar to the problem of setting up lease buyout value in the very beginning. I thought that Tesla bet on increasing volumes swallowing any potential pricing errors into a noise, and that turned out to be true. No matter all the hand wrangling and thousands of SA articles at the time...
There's exactly zero chance that regulators will allow FSD on public highways as anything more than a driver assistance feature without computer-level redundancy, nor is there any chance that Tesla would deploy it that way. The risk of catastrophic failure resulting in death is simply way too high for the automaker to assume responsibility in the event of an accident.
I don't think anybody paid money for a feature called "full self driving" with the assumption that it would never exceed L2 autonomy. Just saying.
Daniel in SD
(supervised)
Just added one!
FSD Beta Videos (and questions for FSD Beta drivers)
There are many other possible reasons. Main one is probably him not remembering accurately and the situations were not identical. But it could also be that Tesla has decided to enable feature that previous were disable as they didn’t have enough validation. Or something else. Here is the exact...
teslamotorsclub.com
Knightshade
Well-Known Member
FWIW this has nothing to do with promises.
An (actual) self driving system that loses the ability to drive if a single node kernel faults or otherwise fails is a serious danger and would likely be banned from public roads pretty quickly, even in the places it's technically legal today- and never approved in the first place in the places it's not.
Similar threads
