Self driving update: Gen 2 and Gen 3 Autonomy on R2 should be similar for a "couple of years"

[macb00kemdanno]

You're the not fun @DuoRivians , huh? I always confuse y'all. One is playful, the other gets butthurt. And it depends on the subject which is which.

Either way, we love you! Time heals all colonoscopies.

Man, I love this place. I don't come to online discussions to be serious all the time (I have enough of that to contend with at work). I love it when we're all loose and free-wheeling; it just makes for a more entertaining discussion.

Sponsored

 

[portdirect]

While we are on this topic, also sharing TOPS comparison with Tesla HW versions: (from Google gemini after couple of prompting):

And the Media Control Unit (Entertainment processor on both platforms): MCUs/Media processors are used for AI Assistant, navigation, music etc.

Re ADAS compute:

• Gen1 had no nvidia.
• Gen2 peak TOPs is around 508 (2x Nvidia Orin AGX at 256 each).
  • Rivian's marketing has muddied the waters here a bit - mostly referring to the capability of a single orin, contrasting that to a pair of RAP1s.

Re infotainment:

• Rivian has never used an Intel Atom - the R1 and R2 use the same SOC for infotainment.

Bad AI.

Edit: I should have said R1 G1 and G2 use the same SOC for infotainment - R2 is on an updated Snapdragon Cockpit SOC

 

[NinjaWrap]

Rivian really can’t afford to create a situation right now that introduces an Osborne effect.

We already saw something similar with Tesla. A lot of owners held off waiting for HW5, only to find out HW4 is likely here to stay for consumer vehicles because Tesla considers it more than capable for their autonomy goals. Meanwhile, HW5 appears to be focused more toward Optimus, robotaxis, and other future platforms.

The Model YL is another good example. Any official announcement about a North American release too early could potentially hurt current Model Y sales from buyers wanting more space.

With Rivian, the situation is even more delicate because their autonomy stack still isn’t as mature. They’re improving quickly, but they’re also clearly behind Tesla in that area. So for people constantly chasing the “next big thing,” they may end up waiting forever instead of enjoying what’s available today and what Rivian already does extremely well.

At the end of the day, there will always be waves of early adopters who buy immediately, others who prefer to wait, and people who stay on the sidelines chasing whatever comes next.

If your priority is the best autonomy experience today, buy a Tesla. If you want the best alternative with an amazing overall ownership experience, buy a Rivian and enjoy the journey.

it might create an Osborne effect for some buyers, but you have to remember only 15% of Tesla owners pay for FSD, despite it being EXTREMELY good now, and that's even after getting free trials when you buy the car. Most people just don't care about autonomy. I wish I didn't!

 

[Electron]

it might create an Osborne effect for some buyers, but you have to remember only 15% of Tesla owners pay for FSD, despite it being EXTREMELY good now, and that's even after getting free trials when you buy the car. Most people just don't care about autonomy. I wish I didn't!

True, however, Tesla is also in a unique position because they’re so far ahead in autonomy right now. FSD has become one of Tesla’s biggest differentiators, to the point where they can remove basic Autopilot from new vehicles and still continue selling cars at scale (much less now, but still way more than other EVs) while also pushing more owners toward the subscription model. $$$$$

That’s something only Tesla can really get away with at the moment. (BMW tried with unlocking heated seats etc...hiding features behind a pay-wall and failed). Whether autonomy personally matters to someone or not, it’s still a major metric among automakers and something every EV (and to a lesser degree, ICE) company has to keep developing to stay competitive.

And while only a smaller percentage of owners may actually pay for FSD, I think cost has always been the biggest barrier. A lot of people simply couldn’t justify the full upfront purchase price (recently ~$8K), but a monthly subscription lowers that barrier significantly. It’s recurring revenue, and Tesla clearly understands the long-term value in that model, and I’m sure Rivian does too.

Rivian is still in the earlier stages of this journey. A $2,500 buy-in for Autonomy+ today feels similar to where Tesla once started before FSD evolved into what it is now. The difference is Rivian now has the advantage of watching Tesla’s roadmap and mistakes in real time, which could allow them to progress much faster than people expect.

And once Rivian’s system becomes genuinely impressive, I think we’ll absolutely see pricing increase and subscriptions become a much larger revenue driver for them too.

The biggest thing people underestimate is the first time someone actually experiences a truly capable autonomy system like FSD. Especially for non-EV people, it completely changes their perspective. Every person I’ve shown FSD to has been speechless afterward. It’s one of those rare tech moments that feels like jumping from a pager straight to a modern smartphone. Once you experience it, your expectations change overnight.

If one can swing it, I think buying Autonomy+ now for $2,500 is a steal. Fingers crossed that Rivian allows transfers for Autonomy+.

 

[hsctiger]

Blue Steel was in the shop for a little more than a week (took them 4 hours to find a dirty connector that was shutting the car down- at least that was the only problem) and I got a loaner with FSD. Won't get another Tesla (R2 reservation) but the feature was really nice.
I assume whatever Rivian comes out with on R2 won't approach this, but I'm cool with whatever because as several of you have already pointed out... it'll get better. And I agree $2500 for this feature is a very nice price for what amounts to better cruise control for now that may ultimately turn into FSD.

 

[Dark-Fx]

<insert long and grumpy early R1 adopter response here>

You should never buy anything that's technology related if you want it to be relevant in more than a year or two.

 

[sparked]

Re infotainment:

• Rivian has never used an Intel Atom - the R1 and R2 use the same SOC for infotainment.

Bad AI.

R2 infotainment is actually 1 generation ahead of R1.

I posted the quotes in a new thread just now to get the word out: https://www.rivianforums.com/forum/...-models-and-higher-performance-than-r1.59467/

 

[shandering]

There are many cars consumer in china with self driving solutions (Xpeng, BYD, xiaomi, Huawei ADS, etc.) and there is also Wayve which is developing an end to end neural network based solution.

What all of these systems have in common is there is zero evidence (literally zero) that adding lidar/radar or more sensors adds any additional safety. Xpeng removed lidar and they claim that transformer neural networks do not understand lindar inputs.

There is also zero evidence of any of them being remotely close to the level of long tail safety that is seen with FSD v14. You can see a video with an XPENG VLA system running over a child and having extremely slow reaction times overall. You can see videos of Huawei ADS systems running into semi trucks on the highway in broad daylight. Pretty sure tesla solved that issue almost 4 years ago.

If you compare reaction time, tesla has around 60ms reaction time

When you see a demo drive with a chinese system and they have an intervention within a couple miles, that's not a good sign. If you use FSD you can drive around in NYC all day and not see an intervention. Using a predetermined route you wouldn't see an intervention , ever.

People like to feel good about themselves claiming that it is the lack of using lidar/radar that is the reason that tesla does not have unsupervised driving on consumer, but the truth is, driving with neural networks and HD maps, no one else is remotely close to doing it. So far that you wouldn't even think it is possible.

There was a teardown of the RAP1 computer from jerryrig everything and it shows that the RAP1 computer has about 36gb of memory. 18gb per chip. Of course the new chip could have more. The memory bandwidth of the RAP1 computer is about 1/2 of what is found in the HW4 computer from tesla.

The reality is, if memory bandwidth is a limitation, then the RAP1 computer may not be suitable for self driving. If camera cleaning is required (as seen in tesla robotaxi vehicles) then no current consumer car will ever be self driving.

If better communications are required, then a car without something like starlink will never self drive. And even that is no guarantee.

What is also unknown is what is the issue achieving the long tail safety that tesla has? Is it some proprietary methods that are unknown, is it a training compute issue, is it a data collection issue, or is it something else?


As far as rivian's claims of Eyes-off self driving next year. That is 100% possible. Level 3 systems are not difficult to build. Building a system with a restricted ODD is not hard and does not require the use of lidar/radar. The reason that manufacturers like Tesla never pursue level 3 (they could have had this tech in 2022) is because it doesn't make sense.

There is zero advantage of a level 3 system over a high level of safety L2 system. L3 systems have a restricted ODD, have to get certified in every state that they operate, and the driver still needs to be paying attention for both taking over when requested and identifying the car having failures or mechanical issues (where the manufacturer is not responsible for accidents). It is a legal gray area whether you can really read a book or watch a movie.

There is no evidence that rivian can do anything that the chinese are not already doing. They have cars with more sensors, more compute than rivian and they have 1000s of employees that have been working on this problem for years

Tesla shows the sobering reality that self driving is not easy and it's not sensors that are the issue. Consumer self driving is many years away.

The other reality is that it costs billions of dollars a year to have all the infrastructure for a self driving solution. It's not going be the case in the short term that there are multiple car companies with this solution.

 

[NinjaWrap]

True, however, Tesla is also in a unique position because they’re so far ahead in autonomy right now. FSD has become one of Tesla’s biggest differentiators, to the point where they can remove basic Autopilot from new vehicles and still continue selling cars at scale (much less now, but still way more than other EVs) while also pushing more owners toward the subscription model. $$$$$

That’s something only Tesla can really get away with at the moment. (BMW tried with unlocking heated seats etc...hiding features behind a pay-wall and failed). Whether autonomy personally matters to someone or not, it’s still a major metric among automakers and something every EV (and to a lesser degree, ICE) company has to keep developing to stay competitive.

And while only a smaller percentage of owners may actually pay for FSD, I think cost has always been the biggest barrier. A lot of people simply couldn’t justify the full upfront purchase price (recently ~$8K), but a monthly subscription lowers that barrier significantly. It’s recurring revenue, and Tesla clearly understands the long-term value in that model, and I’m sure Rivian does too.

Rivian is still in the earlier stages of this journey. A $2,500 buy-in for Autonomy+ today feels similar to where Tesla once started before FSD evolved into what it is now. The difference is Rivian now has the advantage of watching Tesla’s roadmap and mistakes in real time, which could allow them to progress much faster than people expect.

And once Rivian’s system becomes genuinely impressive, I think we’ll absolutely see pricing increase and subscriptions become a much larger revenue driver for them too.

The biggest thing people underestimate is the first time someone actually experiences a truly capable autonomy system like FSD. Especially for non-EV people, it completely changes their perspective. Every person I’ve shown FSD to has been speechless afterward. It’s one of those rare tech moments that feels like jumping from a pager straight to a modern smartphone. Once you experience it, your expectations change overnight.

If one can swing it, I think buying Autonomy+ now for $2,500 is a steal. Fingers crossed that Rivian allows transfers for Autonomy+.

I’ve been saying for years that Tesla should make FSD $49/mo if they want mass adoption. I don’t care how good it is, most people don’t want another $100 on top of their car payment if they’re already stretching to get the car itself.

 

[portdirect]

R2 infotainment is actually 1 generation ahead of R1.

I posted the quotes in a new thread just now to get the word out: https://www.rivianforums.com/forum/...igher-infotainment-performance-than-r1.59463/

Sorry - I meant R1 G1 and G2

 

[shandering]

If we learn anything from Tesla it’s the compute power that is the issue as time moves along.

If we learn anything from tesla than the RAP1 comptuer is not sufficient. It has 36gb of memory and about half the memory bandwidth of the HW4 computer.

The tesla HW4 computer has 32gb of memory and 48gb in the rumored HW4.5 computer in cars right now.

It's clear that memory/memory bandwidth and not TOPS is the issue.

The current planned setup covers most people to strong L3 autonomy and updates for many years.

There is no such thing as strong L3.

What people confuse about autonomy is the levels of autonomy are not performance metrics. L4 is not always better than L3, which is not always better than L2, etc.

L3 is a restricted software with a very limited ODD where the consumer has to both be ready to take over at the split second the car requests it, and also identify any mechanical issues the car is having and take over. While you do not have to look at the road you do have to be paying attention.

Under these restricted conditions (so far it's low speed, no sunrise/sunset/night, lead car, no construction, good lane markings, premapped roads, no emergency vehicles) it's effectively impossible to crash. Tesla could have had this technology in 2022. That means the only crashes are going to be not at fault (other cars hit you), or mechanical failures (suspension, tires, something else) where the driver is at fault.

Because of this L3 is a legally stupid system where you are not really supposed to do anything (other than not having to look at the road). Because doing anything would make you too distracted to correctly fulfill the duties of a L3 system user.

Level 3 technology also does not scale to Level 4 in any way, requires development and certification in every region that you operate and does not financially make sense unless you want to mislead consumers into thinking that level 3 is better than level 2. A level 2 system with high long tail safety, i.e FSD v14 is better than any level 3 system. If you have a strong level 3 system, then you don't have the technology for level 3. You have the technology for level 4.

You can have a level 4 system that only works on highways. The rules of a level 4 system is there are no handoffs to a driver. That means any time the ODD is l eft, the car needs to pull off to the side of the road. The system can only be engaged when on the highway, the user needs to disengage to exit the highway, and if the ODD is not met (highway ending) the car needs to pull over.

 

[Jeremy3292]

If we learn anything from tesla than the RAP1 comptuer is not sufficient. It has 36gb of memory and about half the memory bandwidth of the HW4 computer.

The tesla HW4 computer has 32gb of memory and 48gb in the rumored HW4.5 computer in cars right now.

It's clear that memory/memory bandwidth and not TOPS is the issue.

Where did you find the RAP1 RAM specs? I've yet to see anything published on that front.

 

[shandering]

Where did you find the RAP1 RAM specs? I've yet to see anything published on that front.

The memory bandwidth was stated during autonomy day. It is very low. The memory amount was in the teardown video from jerryrigeverything

Presumably memory is expensive and rivian uses cheaper memory instead of the normal GPU memory that tesla (GDDR6x) is using

The new AI5 computer has something like 192gb of memory. There's a reason it's not going in a consumer car

memory bandwidth requirements would also be higher if you have more sensors.

FSD reaction times are very fast. Around 60ms whereas the chinese systems have extremely slow/nonexistent reaction. You can see tests where users jump out from behind a parked car to see how fast the car brakes. Tesla is fast and the other cars don't react at all. There's a reason other brands can't drive around unsupervised, let alone with someone in the passenger's seat.

 

[Jeremy3292]

The memory bandwidth was stated during autonomy day. It is very low. The memory amount was in the teardown video from jerryrigeverything

Presumably memory is expensive and rivian uses cheaper memory instead of the normal GPU memory that tesla is using

I watched that video many times before and he never says how much GB of RAM. I've been waiting to find the GB specs since Rivian Autonomy day. He only mentioned the 205 GB/s bandwidth which Rivian previously announced already. So we still don't know how much GB of RAM there is.

 

[shandering]

I watched that video many times before and he never says how much GB of RAM. I've been waiting to find the GB specs since Rivian Autonomy day. He only mentioned the 205 GB/s bandwidth which Rivian previously announced already. So we still don't know how much GB of RAM there is.

He tears the chip apart. It's as simple as looking at how much memory there is.

It's literally right there. 18gb

Keep in mind tesla has 384 GB/s in a chip from 2022. 205GB/s is ridiculously low

Sponsored