-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
HW4 & HW5 Discussion
If this brings back memories of the 2016 FSD drive on YouTube, there should be no doubt this is different. It is exactly why Elon did this live vs during some other "staged" event.
I did consider this could be Palo Alto data rich. But that just means more compute would be needed to build it out for more scenarios and that we should see this progress everywhere fairly soon.
Failing and changing the approach is how progress is made.
Elon thought success was right around the corner. This is a trait of great inventors. It keeps them motivated.
Another trait is that they keep trying after they fail. Yet another is that they learn from their mistakes.
I've thought for a long time that FSD training was compute-constrained. Now that computing power is being ramped up by orders of magnitude, I too think success is right around the corner. I could be wrong, but it won't be because Tesla failed in the past.
"Past performance is not indicative of future results." That old investing maxim goes for both success and failure.
The "Chat GPT moment" refers to the time when investors start to believe that huge profits are imminent. It does not refer to the point where the technology is perfected. So no, we are not there yet. When Tesla stock, driven by the FSD narrative, spikes even sharper than in 2020 you will know we are there. That's how you will know that the "Chat GPT moment" has arrived.
Bumping this tweet by the guy in charge of Tesla training infrastructure:
Raw photons are easier to count because they are firmer and less mushy than after you grill, bake, braise, or otherwise cook them.
I watched a good bit of the Elon driving video. It is noteworthy to me that he commented that the car didn't need an ongoing internet connection for FSD to keep working. It seems like with very little work, you could have a Tesla FSD take basic directions from the driver/passenger, like "turn left here", "take the second right after the stop sign", etc., as you would if you were talking to a cabbie who didn't know where your destination was. Perhaps this would finally be the JohnnieCab from Total Recall?
Not sure how important/useful that would be, since seems likely every car would have a basic map, but it could work if the driver/passenger didn't know a street address for where they are going, or perhaps some other edge cases. Plus allowing the common desperate chase scene action hero line "Just drive!!!"
Not sure how important/useful that would be, since seems likely every car would have a basic map, but it could work if the driver/passenger didn't know a street address for where they are going, or perhaps some other edge cases. Plus allowing the common desperate chase scene action hero line "Just drive!!!"
Thanks for all the info. It made me consider the question... will I want an H100 type system anytime soon for the home? (Recalling a similar situation decades ago from the late Steve Jobs). The answer is likely OT, and pretty far into the future as far as "Home Learning" goes.
Todd Burch
14-Year Member
How many megawatts does your home panel support?
Last edited:
Or looking for your cat, dog, grandpa, or even a child based on a photo you put in front of the camera and press "Find".
All a part of the untapped revenue stream of a brain on wheels. We can't even imagine all the Apps possible.
Silly question on several levels. This is typical of what they'd say to Steve Jobs regarding the PC by the way.How many megawatts does your home panel support?
I don't need a Farm, maybe a dozen, IDK. Have you looked into it? I want to teach my robot how to maintain my backyard - I'm getting old.
From Nvidia site: "The H100 PCIe Gen 5 configuration provides all the capabilities of H100 SXM5 GPUs in just 350 Watts of Thermal Design Power (TDP)"
Last edited:
Knightshade
Well-Known Member
Based on physical teardowns of HW4 it appears to be exactly the opposite of that.
Virtually 0 architectural difference, just more/faster cores.... (CPU cores go from 12 to 20, and TRIP cores go from 2 to 3)
Note the below is on an S/X HW4 computer- the ones from the Y appear to be lesser HW (we covered this a while ago though- less storage/memory on the infotainment part- and less camera connectors populated on the driving computer part)
There are many people messing around with ML using gaming GPUs from Nvidia and AMD. No server grade GPUs needed(plus they are terrible to work with).
Really hard to tell just by looking at it. Usually we see a pretty big IPC uplift from first gen to 2nd due the high amount of low hanging fruit that gets optimized. So the chip can be 10%-200% faster core v core running at the same frequency depending on how the chip is fed memory bandwidth wise(which wasn't specified from green's picture.
I think I already saw that in ‘The Terminator’. Works well.
2daMoon
Mostly Harmless
It is obvious how posting something FSD related would be off-topic for the investment forum , but, I just had to share Lars' latest video that offers a few pieces of data about "the competition" and their supply chain that seemed interesting.
, but, I just had to share Lars' latest video that offers a few pieces of data about "the competition" and their supply chain that seemed interesting.
We know a handful of HW4 owners have FSD beta V11 and are testing it.
So if special training is needed to support V11 on HW4 Tesla as done it at least once and it is reasonable to conclude they could do it again.
Initially the Tesla Semis didn't have FSD, that is probably because they are HW4.
My guess is Tesla is trying to get a version of V11 that is suitable for wider release if only because that is the initial version HW4 owners will get.
Most drivers testing FSD beta are testing V11, it is better than older version of FSD and not a bad initial version to get.
V12 is alpha outside of Tesla not many people are testing it, Elon now considers V12 to be FSD V11 is ultimately a dead end.
My conclusion is that for V12 it will work best on HW4 with specific compute intensive training for HW4.
IMO it is very likely that HW5 will have the same sensor suite as HW4 and there wil be an upgrade path from HW4 to HW5.
As a signal processing and machine learning engineer, I had been waiting for the day Tesla would stop using the former (ad-hoc signal processing code) and rely only on the latter.
Literally since 2017: [Speculation] Tesla going to use End-to-End Deep Learning to Control Vehicles
The reason, of course, is that while rigid code can work on a simplistic problem or one with low thresholds for accuracy, it is terrible to scale when a problem is complex or high accuracy needed, let alone both. Even earlier this year as people were excited with V11 release, I said "wake me up when we are using all neural nets". Well finally that time has arrived.
Tesla has now turned the problem from engineers focused on debugging their indivdual models to a focus on data cleaning. With an architecture that already shows promise in driving smoother (allegedly) than V11, decisions on the exact data being fed in becomes really important.
There is still going to be many interations, becuase each time the model runs, they will identify the current failure modes and work to secure the additional (or reweighting) of data that best exemplifies those edge cases and how they want the model to perform.
While this may be trivial at first, I do believe it takes more skill each successive training.
In addition, just because Tesla is fully neural nets does not mean this architecture will be the last. They will still be working on iteratively improving the architecture. But working one 1 architecture vs say 52 independents modules is a heck of a lot less engineering labor time.
So I tentatively believe a pretty advanced L2 system will come from V12 sometime early next year. Robotaxi level is still hard to estimate simply because of accuracy thresholds needed - there still may need to be a lot of iterations to get it to a useable point.
But at least now we can see the light at the end of the tunnel.
Literally since 2017: [Speculation] Tesla going to use End-to-End Deep Learning to Control Vehicles
The reason, of course, is that while rigid code can work on a simplistic problem or one with low thresholds for accuracy, it is terrible to scale when a problem is complex or high accuracy needed, let alone both. Even earlier this year as people were excited with V11 release, I said "wake me up when we are using all neural nets". Well finally that time has arrived.
Tesla has now turned the problem from engineers focused on debugging their indivdual models to a focus on data cleaning. With an architecture that already shows promise in driving smoother (allegedly) than V11, decisions on the exact data being fed in becomes really important.
There is still going to be many interations, becuase each time the model runs, they will identify the current failure modes and work to secure the additional (or reweighting) of data that best exemplifies those edge cases and how they want the model to perform.
While this may be trivial at first, I do believe it takes more skill each successive training.
In addition, just because Tesla is fully neural nets does not mean this architecture will be the last. They will still be working on iteratively improving the architecture. But working one 1 architecture vs say 52 independents modules is a heck of a lot less engineering labor time.
So I tentatively believe a pretty advanced L2 system will come from V12 sometime early next year. Robotaxi level is still hard to estimate simply because of accuracy thresholds needed - there still may need to be a lot of iterations to get it to a useable point.
But at least now we can see the light at the end of the tunnel.
Waymo and Cruise already have functional robotaxis, and you're saying it's still going to take Tesla years to get there? Huge lead Cruise / Waymo can build in major metros.
Similar threads
