-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Is it smart to buy FSD for $5k while it’s still available for MX owners?
- Thread starter rjdoc74
- Start date
I bought FSD not long ago just to support our favorite uncle Elon. IMHO he deserves that just for doing such a great job designing my amazing 2018 Model X75 D. That said, I was expecting just a teeny bit more out of ver. 9 than I thought it delivered, and that fact correspondingly lowered my FSD expectations.
bayx
Member
Cooler heads always prevail!Most sensible!
I don’t know. If it’s something as cool as self driving it may support a significant premium at least for a while.
lux_cars
Member
You’re forgetting competition. Tesla can’t charge way more than competitors will for self driving. They might even get there first.
Actually, v9 is turning out to be better than I expected. I accidentally left mine in autosteer in a turn lane, and the darn thing followed the car in front of me around the corner. This might have been a fluke.
Interpreting stop lights and stop signs is actually relatively easy. The reason they aren't doing it yet is that every time you add a new image recognizer model (e.g. for road signs), it takes more processing power, and the existing hardware can't even handle full-frame-rate video without those extra models.
The current hardware can barely manage what is being thrown at it. Spending man hours trying to squeeze a few more frames per second out of their model processing code so that they can add that feature on existing dead-end hardware is not a good a use of their engineers' time. Better to spend that effort developing those features (and more) on the hardware that will actually be running it. Save the performance tuning for when the models start to push the limits of AP3.
To put things into perspective, right now, as I understand it, the AP2.5 hardware can run their current machine vision models at only 200 frames per second (and considerably fewer on the slower AP2 hardware). Divide that by 8 cameras, and that gives you only 25 frames per second per camera. The cameras provide 60 frames per second worth of video. This means that the AP hardware is either doing low-res processing on some cameras, ignoring some cameras entirely, or ignoring more than half of the frames of video coming in from every camera (and sometimes two frames in a row) just to get by.
BTW, this is, I'm quite certain, why AP2 hardware doesn't support the dash cam feature. If it ran the front camera at 60 FPS, it would not have enough processing horsepower to do everything else that it needs to do. It isn't that it isn't quite powerful enough, but rather that it isn't anywhere NEAR powerful enough. Even with AP 2.5, it only manages 36 fps.
And prior to the major rewrite a few months ago, those machine vision models ran even slower. My best guess estimate, judging by how much latency it had when dealing with changes to the curvature of the road, was that it was processing maybe five or six frames per second per camera. The odds of it happening to notice a street sign are abysmal at that rate, and there's no way you could add the extra processing on top of what it was already doing. That extra processing would have a decent chance of causing a wreck. However, the rewrite put their machine vision system within the realm of possibility when combined with a hardware upgrade.
The new AP3 hardware can run their existing models at 2,000 frames per second, which is 250 frames per second per camera — more than 4x their full data rate. That means that they'll be able to increase the complexity of their models by more than 4x before being forced to find ways to optimize them further. That's hopefully enough growing room for all the secondary models that they'll need to add, like the model for interpreting text on signs after the main recognizer detects them, interpreting traffic lights, determining the direction of motion for vehicles and pedestrians up ahead, turn signal interpretation, driver and cyclist hand signal interpretation, various other intent prediction models, etc. that all have to be run on top of the main image recognition.
As I said earlier, many of those pieces (not all) are relatively straightforward, and no doubt have been under development independent of the main machine vision model and the software and hardware used to run those models. But until you have enough CPU/GPU/TPU horsepower to run the existing models at a usable speed, plugging those extra pieces in would be a total non-starter.
Given that Musk said FSD features would start rolling out in version 9, and that Tesla is beta testing FSD on employee vehicles, I wouldn't expect #1 to be the case.
Of those, my guess is that it is a combination of the other three.
The hardware updates probably aren't significantly more than $3k. But the Model 3 is selling in volume, so small amounts of money add up to big losses that they probably don't want to absorb, particularly given that they think the feature will sell itself to existing customers even at full price (#4) once it comes out. The reason it is still available as an upgrade from your account at full price ($5k) tells us that the cost of the HW3 MCU is likely between $3k and $5k.
That said, I think there are two other factors that you missed that probably play a bigger role:
- Priority upgrades for Model S and Model X. When they start rolling out upgraded hardware, it's going to be a huge upgrade effort for a lot of vehicles. Every extra vehicle with prepaid FSD means that the upgrade process will take longer. Most new sales will be for the new, lower-cost model of the Model 3. By temporarily stopping sales of FSD, Model S, Model X, and high-end Model 3 cars will get their upgrades sooner than if they make the FSD feature available to the giant horde of Model 3 owners buying in at the newly opened next price tier down.
- The AP3 hardware will start rolling out in the factory soon-ish, as in early next year. Many cars ordered now won't be delivered until then, which means if they make the FSD feature available, most of those cars would end up needing a new MCU before they even leave the factory, and they would have to figure out which cars need which MCUs based on their configuration so that they can clear out their stock of existing MCUs. By forcing the upgrade to be deferred until a post-delivery servicing, they keep installing AP2.5 MCUs at the factory until they run out, and then the repair centers can just store their working pulls as spare parts for future AP2.5 repairs. That's a lot easier to deal with on a one-off basis at a service center than at the factory.
Actually, v9 is turning out to be better than I expected. I accidentally left mine in autosteer in a turn lane, and the darn thing followed the car in front of me around the corner. This might have been a fluke.
Interpreting stop lights and stop signs is actually relatively easy. The reason they aren't doing it yet is that every time you add a new image recognizer model (e.g. for road signs), it takes more processing power, and the existing hardware can't even handle full-frame-rate video without those extra models.
The current hardware can barely manage what is being thrown at it. Spending man hours trying to squeeze a few more frames per second out of their model processing code so that they can add that feature on existing dead-end hardware is not a good a use of their engineers' time. Better to spend that effort developing those features (and more) on the hardware that will actually be running it. Save the performance tuning for when the models start to push the limits of AP3.
To put things into perspective, right now, as I understand it, the AP2.5 hardware can run their current machine vision models at only 200 frames per second (and considerably fewer on the slower AP2 hardware). Divide that by 8 cameras, and that gives you only 25 frames per second per camera. The cameras provide 60 frames per second worth of video. This means that the AP hardware is either doing low-res processing on some cameras, ignoring some cameras entirely, or ignoring more than half of the frames of video coming in from every camera (and sometimes two frames in a row) just to get by.
BTW, this is, I'm quite certain, why AP2 hardware doesn't support the dash cam feature. If it ran the front camera at 60 FPS, it would not have enough processing horsepower to do everything else that it needs to do. It isn't that it isn't quite powerful enough, but rather that it isn't anywhere NEAR powerful enough. Even with AP 2.5, it only manages 36 fps.
And prior to the major rewrite a few months ago, those machine vision models ran even slower. My best guess estimate, judging by how much latency it had when dealing with changes to the curvature of the road, was that it was processing maybe five or six frames per second per camera. The odds of it happening to notice a street sign are abysmal at that rate, and there's no way you could add the extra processing on top of what it was already doing. That extra processing would have a decent chance of causing a wreck. However, the rewrite put their machine vision system within the realm of possibility when combined with a hardware upgrade.
The new AP3 hardware can run their existing models at 2,000 frames per second, which is 250 frames per second per camera — more than 4x their full data rate. That means that they'll be able to increase the complexity of their models by more than 4x before being forced to find ways to optimize them further. That's hopefully enough growing room for all the secondary models that they'll need to add, like the model for interpreting text on signs after the main recognizer detects them, interpreting traffic lights, determining the direction of motion for vehicles and pedestrians up ahead, turn signal interpretation, driver and cyclist hand signal interpretation, various other intent prediction models, etc. that all have to be run on top of the main image recognition.
As I said earlier, many of those pieces (not all) are relatively straightforward, and no doubt have been under development independent of the main machine vision model and the software and hardware used to run those models. But until you have enough CPU/GPU/TPU horsepower to run the existing models at a usable speed, plugging those extra pieces in would be a total non-starter.
very interesting read, but it also supports my theory that FSD is still years away given all they have to implement. and at least 1 processor upgrade if not more away, depending on what they find once they start implementing all those things you listed. my bet is 3 years away whereas Waymo will be out next year... but in a commercial version and not individually operated.
What Tesla should do is rebrand the FSD option (to EAP+?) and provide a better description of what you are getting when you purchase it.
When we purchased FSD for our S and X, our assumption was that even if Tesla doesn't get the software approved for full self driving, that the software would drive the vehicle under increasingly more conditions, in driver assist mode.
With V9, once the navigate on autopilot is enable for entry ramp to exit ramp, it appears Tesla is getting close to delivering on the goals for EAP.
If EAP+ replaced FSD, and was defined as providing features beyond EAP, in driver assist mode initially, and potentially as full self driving in the future (upon completion of validation testing and receiving regulatory approval), then owners should start seeing benefits of owning EAP+ soon (later in V9, or in V10). Plus, for customers who've purchased EAP+ (including those who already bought FSD), they would be provided free upgrades to the AP processor, when needed to support the full features of EAP+ or to achieve FSD.
Providing an EAP+ option, with the potential for future FSD, provides Tesla a competitive edge over the other manufacturers, who are more likely to provide only the capabilities they already have working in the vehicles they sell.
Even if Tesla never got FSD approved for use, our purchase of the FSD option would be justified, if Tesla can provide us "navigate on autopilot" for most of our driving, in driver assist mode - not just on limited access highways...
When we purchased FSD for our S and X, our assumption was that even if Tesla doesn't get the software approved for full self driving, that the software would drive the vehicle under increasingly more conditions, in driver assist mode.
With V9, once the navigate on autopilot is enable for entry ramp to exit ramp, it appears Tesla is getting close to delivering on the goals for EAP.
If EAP+ replaced FSD, and was defined as providing features beyond EAP, in driver assist mode initially, and potentially as full self driving in the future (upon completion of validation testing and receiving regulatory approval), then owners should start seeing benefits of owning EAP+ soon (later in V9, or in V10). Plus, for customers who've purchased EAP+ (including those who already bought FSD), they would be provided free upgrades to the AP processor, when needed to support the full features of EAP+ or to achieve FSD.
Providing an EAP+ option, with the potential for future FSD, provides Tesla a competitive edge over the other manufacturers, who are more likely to provide only the capabilities they already have working in the vehicles they sell.
Even if Tesla never got FSD approved for use, our purchase of the FSD option would be justified, if Tesla can provide us "navigate on autopilot" for most of our driving, in driver assist mode - not just on limited access highways...
We have an MX order in flight and added it to the purchase after-the-fact for $3k. Our purchase will be fully deductible via IRS 179, so figure the net cost is more like $2k. I can't reasonably apply the same deduction to the $5k future upgrade pricing, so this was a no-brainer.
My belief is there will be incremental features rolled out to vehicles with HW3 (similar to HW2) and at 3% of the cost of the vehicle, it's a reasonable investment.
My belief is there will be incremental features rolled out to vehicles with HW3 (similar to HW2) and at 3% of the cost of the vehicle, it's a reasonable investment.
Definitely. Tesla’s been selling FSD for 2 years without issue to S/X owners and now more 3 owners show up it’s causing too much confusion.3 owners are way more demanding, cynical, and impatient than X owners.
The practice was dick-ish. You bought the hardware but are artificially limited to what you can use.
No different than having a large gas tank but having its capacity limited until you pay extra to have it unlocked. Or buying a house and one of the rooms was locked off until you paid an additional 10k.
Actually, it's worse than that -- the person who bought the 60 had to lug around the additional, unusable weight of the 15 kWh worth of battery.
No wonder they abandoned the practice.
mdk1a
Member
There's the question of purchase timing (buying it now to potentially save vs later), but if you do it now, I think there's also the question of what exactly does it include and when will you actually get it. For example, if your intention is to really buy "FSD" now, doing so may include HW3 for free whenever it comes out, but I'm not sure that HW3 is necessarily = FSD. If additional HW upgrades beyond that are going to be needed to really achieve FSD, are they included if you buy the FSD package now? It would seem to me that HW upgrades are way costlier than SW ones, and I think it would be foolish for any company to say, "hey, pay us $X amount now, and we will *guarantee* that we will do *all* the necessary HW upgrades and for however long it takes at no additional cost to you to achieve some feature Y". I don't believe Tesla has said anything to that effect either.
Personally, under the current umbrella "FSD" term, I think there's too much ambiguity around just exactly what it includes and by when, and the whole thing also seems too much about the potential. When (hopefully it's a when question) Tesla really comes out with an actual FSD feature, I will be more than happy to pay them whatever they deem to be the right market price is. Heck, depending on what other upgrades/new features come with the latest model at the time, I will highly consider trading-in my current one too.
Personally, under the current umbrella "FSD" term, I think there's too much ambiguity around just exactly what it includes and by when, and the whole thing also seems too much about the potential. When (hopefully it's a when question) Tesla really comes out with an actual FSD feature, I will be more than happy to pay them whatever they deem to be the right market price is. Heck, depending on what other upgrades/new features come with the latest model at the time, I will highly consider trading-in my current one too.
Here's the strategy:
2016: FSD won't be ready in at least 5 years
- Sell cars with FSD for 2 years
- Offer free HW upgrade if needed
2018: FSD won't be ready in at least 3 years
- Stop selling FSD to accumulate more obligations we can't fulfill
2019: 3 year leases ending, people start trading in for cars with HW3+
- Disable FSD feature on all incoming inventory/CPO AP2/2.5 cars
- Push off people who demand retrofit because FSD software is not ready. Pointless to retrofit for free.
2021: FSD starts rolling out
- We only sold FSD cars up to until 2018. Most of them already moved on/traded in for the HW3+ cars
- Continue to disable FSD feature on all incoming inventory/CPO AP2/2.5 cars
- Retrofit obligation significantly reduced for remaining 2.0/2.5 owners.
Profits on all the AP2/2.5 FSD sold!
2016: FSD won't be ready in at least 5 years
- Sell cars with FSD for 2 years
- Offer free HW upgrade if needed
2018: FSD won't be ready in at least 3 years
- Stop selling FSD to accumulate more obligations we can't fulfill
2019: 3 year leases ending, people start trading in for cars with HW3+
- Disable FSD feature on all incoming inventory/CPO AP2/2.5 cars
- Push off people who demand retrofit because FSD software is not ready. Pointless to retrofit for free.
2021: FSD starts rolling out
- We only sold FSD cars up to until 2018. Most of them already moved on/traded in for the HW3+ cars
- Continue to disable FSD feature on all incoming inventory/CPO AP2/2.5 cars
- Retrofit obligation significantly reduced for remaining 2.0/2.5 owners.
Profits on all the AP2/2.5 FSD sold!
I’m thinking they are doing this to limit their exposure to AP3 hardware upgrades.
Plenty of owners out there that did not select FSD. If the vast majority now decide to add the feature that’s a LOT of cars for Tesla to do hardware upgrades for “free” - hardware installation is going to be a huge cost for them and a logistical nightmare in terms of already overrun SC’s having capacity to do it.
Plenty of owners out there that did not select FSD. If the vast majority now decide to add the feature that’s a LOT of cars for Tesla to do hardware upgrades for “free” - hardware installation is going to be a huge cost for them and a logistical nightmare in terms of already overrun SC’s having capacity to do it.
Similar threads
