Upgrading FSD CPU

Upgrading FSD CPU

I have a late 2016 (AP HW2) S90D and just listened to the Elon - Lex Fridman interview on the MIT AP Driver Awareness study. A couple of things that I got from the interview were:

There are about 400,000 cars with FSD hardware.
The new Tesla CPU for FSD has very recently been added to the cars.
Elon is driving the debug version of the firmware that has most if not all of the FSD functionality.
There is a real chance Tesla will release the FSD functionality this year even if it still requires driver participation.

So what I am wondering is how long it will take for Tesla to upgrade the all CPUs. Does anyone know how many FSD cars out there purchased FSD? Lets say it is half. So 200,000 cars that need a FSD upgrade. Lets say arbitrarily that it will cost Tesla $500 to upgrade a car. (cost of hardware and installation). Unless my math is wrong that would be 100 million dollars. Do we see that as a problem? Tesla has already shown that it prioritizes the current hardware set over the older when it comes to delivering firmware updates. How long before Tesla delivers the FSD to legacy hardware cars after it delivers it to the latest and greatest?

The one factor that could help is how much Tesla has had to modify their earnings because they could not use a portion, possibly none of the FSD sales because they have not delivered any functionality yet.

Anyone have any idea how much that would be? Using the above estimate of 200,000 FSD sales at $2,000/sale that would be 800 million. And how does not being able to recognize revenue effect cash flow?

Any ideas out there? | 18/04/2019

50% FSD seems quite high. My estimate would be closer to 10-20%. Remember many people didn't buy EAP, then many who bought EAP are happy with what it offers. I'm not aware of Tesla giving out the numbers, so we can only guess.

I think your $500 number for Tesla's actual parts and labor costs is sound. Even at 50%, I don't see $100 M in costs a problem. The FSD funds are in reserve and can't be counted as revenue, and while adding HW3.0 without FSD software will not allow complete shifting the money into revenue, I can see Tesla moving perhaps half the value to revenue - so there is a real benefit for Tesla to do HW3 retrofits.

Likely all FSD monies collected is in Tesla's cash, so the retrofits would take a hit to cash. Doubt it will be material compared with vehicle sales.

Also with Tesla's redefinition of FSD for current orders, they may be able to count the revenue long before all FSD features roll out - likely before year end. They only have to deliver two features: Recognize and respond to traffic lights and stop signs and Automatic driving on city streets.

Dramsey | 18/04/2019

"The FSD funds are in reserve and can't be counted as revenue..."

Cite? I wouldn't expect Tesla to sit on the thousands of dollar I paid for FSD back in late 2016, nor do I see why they should. I'd assume that someone at Tesla has run the number on how much it will cost to upgrade the folks like me who prepaid to the new computers, and that would be in the "debit" column, but I really doubt there's a "reserve" fund full of FSD money...

bill | 18/04/2019

As I see to understand it based on google research I believe they can use the money but have a liability on the books for the amount unrecognized. Si it seems it is like a loan without interest.

Also I don't think Tesla can change the definition for the cars it sold already based on a definition at the time. So when I bought in late 2016 FSD meant being able to go from any point a to any point b that you could by manually driving without having to manually drive the car.

bill | 18/04/2019

BTW according to Google "Tesla Inc's current deferred revenue for the quarter that ended in Dec. 2018 was $1,423 Mil."

dougk71 | 18/04/2019

Likely FSD prices will increase a lot starting in May 2019. This will pull forward demand not unlike the tax incentive reduction effective Dec 31 2018. If FSD is delivered at a significant degree of functionality Tesla should be able or may even be required to recognize as revenue all deposits already on the books. For cars with the new hardware all Tesla has to do is push the software for the hardware upgrades if the install is a unplug and plug in new module Tesla could complete the roll out via its service capability by the year end.

pgkevet | 18/04/2019

Whatever functionality is released will likely be 'beta software' requiring driver oversight. I cannot reconcile that with anything called FULL self driving.
The system as is is reactive rather than proactive... there is no anticipation that I recognise (yet?) so when it stops at junctions I'd anticipate it then asking confirmation to proceed and having no ability to cope with complex junctions and frankly there's more of those than simple one's in most of the UK. Polygns with N sides-worth of corner cases...
What does city streets mean? Main thoroughfares only I'd guess.
It has to be baby steps in development but I anticiapte that for many the stress of watching and trying to compensate for computer errors will make it calmer to just drive the thing oneself.

reed_lewis | 19/04/2019

My salesperson told me that he estimated that the 'take rate' for FSD was something in the 2-3% range. Now this was before the recent discounts, but I would not expect that more than 10-15% (totally SWAG) of the cars have FSD provisioned.

Remember the few people here are the fanatics of Tesla, and the average owner does not follow every bit of Tesla news. Among the population here I would guess that the FSD take rate is much higher.

EVRider | 19/04/2019

I agree with @reed that the FSD numbers are probably lower than most of us here think. I recall reading somewhere that the number of owners who buy AP was a lot lower than you would think, but I don’t recall the source or numbers (this was before Tesla introduced “basic” AP). If most owners bought AP, Tesla probably wouldn’t need to mess with the pricing as often as they do. | 19/04/2019

If Tesla is following GAPP accounting (and they say they are), they can't count as revenue money they collect and have not delivered. This includes car deposits, reservation fees, FSD, etc. They can use the money in available cash, but it's still a liability.

I don't think anyone expects FSD to be 100% on day one. It's likely to be in the 80% to start (i.e. handles 80% of point to point driving). It may take another couple of years to get to 95%, and maybe 10 years to get to 99%. These are just my WAG on timing, but it's clearly not going to cover 100% for a very long time, if ever. Still there will be a point when Tesla considers it delivered as far as revenue accounting. This doesn't mean they stop making improvements, but is strictly an accounting issue.

As to what issues will remain unresolved at 80% or 99% - that's yet to be known. My guess is many of these issues will be geo-fenced - meaning it either has to figure an alternate route, or return control to the driver. It might even be an option like - if re-routing to maintain FSD requires more than 5% of the route distance over what than an optimal human route - return control to the driver.

Now there will be arguments what FSD means forever. If it handles 80% of driving is that FSD? Does it have to be 100%? I think we each have our definition, but it really doesn't matter a lot. What matters is what we get when, is it safe, and does it do what we want.

Added to this is what are the alternatives (i.e. competitors) doing. If Tesla continues to be ahead of everyone else, there isn't a lot to complain about. If they fall behind, that could be another story. In these cases, I'm talking about what you can personally buy, not vehicles that can't be purchased. WayMo's $25K ICE vans with a $80K sensor suite/cpu is not something available or comparable.

dougk71 | 19/04/2019

Well, GAPP requires recognition of software and hardware if it is partially delivered. Software delivered to a client that is useful yet only partially delivered is brought onto the books by the supplier. This avoids the issue where a multi-million dollar multi-year software project could be brought onto the supplier's books in one day distorting revenue.

Stiction | 20/04/2019

To me FSD means you really could let the car drive without a driver ....without having a multi K$ prang every year or two. I don't see that happening soon.

With the end of Moore's "law" the cost of the electronics will not go down as fast and nor will the processing power get better at anything near the same pace. All exponentials have to come to an end sometime ..eventually there aren't enough atoms to make a transistor. And yes, there will be advances (eg.. more 3D packaging, less porky software pissing away performance, new display tech) but so far there is NO sign of anything with an exponential curve for computation. More like percents per year.

We take for granted advances in electronics because of the decades of exponential improvement in chip density, speed and power. No longer: The stuff under the tree Christmas morning isn't going to be as nifty, and upgrading your phone to the new model not nearly as compelling.

So color me skeptical that FSD will become functional and cheap anytime soon.

dougk71 | 20/04/2019

Sequential ( regular PC type processing) is limited to cycle speed which in turn is limited to die sizes. Sure PC have multiple CPU's but here are limited situations where true parallel processing is possible.
Now, image processing is very very much different. To some degree pixels can be processed in parallel. With high resolution that means mega pixels can be looked at simultaneously. The this is done by writing code to silicon so electrons can traverse the SOC in parallel effectively processing several thousand pixels at the same time. In a regular PC work is done instruction by instruction and little in parallel.
Tesla wisely has understood the limitations of serial processing and Hardware 3.0 is massively parallel relative to Hardware 2.5. With 3.0 all 8 cameras at full resolution can be processed at once.
A ten fold increase in speed is easily obtained. The downside is that parallel processing logic is unlike traditional processing programming. It is like a traditional code set has every line in a 10,000 line program executing at once instead of line by line. Tesla with HW2.5 is doing much processing line by line resulting in Tesla having to reduce the pixel resolution and go to gray-scale from color cameras just to have calculations complete fast enough for driving.

Dramsey | 20/04/2019

Actually, @dougk71, the Nvidia Drive PX2 computer used in HW2.5 is also a massively parallel processor, with two Pascal-type GPUs, each of which has hundreds of CUDA cores (Nvidia's a little vague about the exact specs on these chips), as well as several ARM cores. You can read a nice overview article here:

"Parallelizing" software is quite difficult, and we don't know anything at all about how Tesla's autopilot software is designed, written, and implemented, so I don't think you have any basis for saying that they're doing a lot of "line by line" processing with the current hardware.

Elon has said that the custom Tesla chip used in HW3 is a much, much faster. Well, Elon says a lot of stuff. I'm sure it''s faster, though. But remember that Elon said for years that HW2 was all that was needed for full autonomy (it was even on their ordering web page, something along the lines of "Every Tesla comes with all the hardware necessary for full self driving capability", before changing his mind and announcing HW3.

dougk71 | 20/04/2019

I understood that HW2.5 contained parallel processing chips and leveraged NVIDIA's Graphic processing expertise. The issue is the Graphic processing for PC type video is very different than object recognition which is needed for FSD. Tesla realized that this NVIDIA parallel solution only goes so far and much linear code was written to get object recognition.I suspect the new hardware does much object recognition directly in silicon and can recognize several objects ( cars people lines etc. ) in parallel providing an order of magnitude improvement in performance and accuracy.

Dramsey | 20/04/2019

The CUDA cores in Nvidia silicon can be used for a broad variety of parallelizable problems, which is why they build supercomputers with them.

When you say "Much linear code was written to get object recognition", it seems you have some very specific knowledge of how Tesla's code works. Is this the case?

Presumably the HW3 chip is optimized for one or more specific tasks related to FSD. Given the speed with which it was developed-- less than 18 months from the announcement in December 2017 to being installed in cars today-- I'd guess that it's based on some existing chip architecture, as opposed to something designed from scratch. Heck, it might even just be a bunch of ARM cores.

Disclaimer: I've been a programmer for over 30 years but have never done any image processing work.

sr.smr | 20/04/2019

All this discussion has me psyched for the Tesla Autonomy Investor Day webcast on Monday.

Dramsey | 21/04/2019

I'm sure it will be fun and impressive. However, years of missed schedules and broken promises have made me very skeptical we'll see any dramatic features added to our cars any time soon. Heck, I (along with many others) am still waiting for 2019.12 and the auto-lane change feature, although others have had it for a month or more. | 21/04/2019

@Dramsey - Others that have .12.x are in the early access program, and it has not be rolled out yet to general owners. It does show that Advanced summon and auto-lane change is coming. Not quite sure why you're so down on AP. I've been getting significant updates and new features for over two years now and looking forward to Advance Summon and auto-lane change - both which seem like dramatic features that no other car maker offers.

Considering there is no vehicle that you can buy other than Tesla that does as much AP. I guess you could buy Audi's answer that only works to 37 mph. GM's supercruise is limited to pre-tracked freeways. Seems everyone else are years behind Tesla. So Tesla has missed a few expected deliveries, it still blows away the alternative solutions. It's not like Tesla is trying to delay or slow down AP features - this is really complex stuff that has never been done before.

EVRider | 23/04/2019

@TeslaTap: 2019.12 is not an early access release, plenty of mere mortals like me have it. No Enhanced Summon in that release. Auto lane change without confirmation during Navigate on Autopilot was already available in 2019.8.5. | 23/04/2019

@EVR - thanks! Good to know, although perhaps you're more than a mere mortal :)

dougk71 | 23/04/2019

Tesla has decided NVIDIA is not the way to go. NVIDIA needs to make chips that many many can use. Tesla is mostly only concerned with FSD. Tesla explained they used existing IP at 14 microns to speed the development of the new architecture that is an order of magnitude faster than the best NVIDIA chip. It is not that NVIDIA couldn't make a dedicated FSD chip it is that it doesn't. Tesla understood that very fast object recognition from a camera frame is essential to FSD. Further they understood that FSD can't be achieved with software or simulation alone. The visual cortex is a neural network and it is not controlled by software. It from birth to death modifies the visual cortex so as to learn from its input. NVIDIA wasn't going to be the solution so Tesla went with its own Network processor that with time with eliminate most of the current software used with NVIDIA chips.
Now the human cortex can't quickly do millions of matrix dot products per second so some mathematical operations will still be software controlled.
They need to mimic the human visual cortex processing only with more visual input ( more than two eyes ).

EVRider | 23/04/2019

@TeslaTap: FYI, TeslaFi's firmware tracker ( shows that 7.6% of the fleet that they monitor has 2019.12. The largest deployment of recent updates is 2019.8.5 at 57.5% (that's what I have on our Model 3).

barrykmd | 23/04/2019

I'm still on 2018.50.6, which is fine by me. Actually, I'd rather be on 2017.x.x.

Dramsey | 24/04/2019


“Others that have .12.x are in the early access program…”

Ah, the program Elon promised I’d be invited to join since I’d purchased FSD more than two years ago. (Checks mailbox). Nope, nothing yet.

“Not quite sure why you're so down on AP.”

Oh, I’m not! I love EAP and use it virtually every time I drive the car. It’s just now at the point I thought I was buying in December 2016. My frustration is with the missed schedules and continual promises that don’t pan out.

“So Tesla has missed a few expected deliveries,”

I think it’s missed pretty much all of them. Since my first Model S in 2013 I don’t recall the company meeting a single self-imposed deadline for, well, anything.

“…it still blows away the alternative solutions.”

This is still true!


I think you need to be more careful when you make definitive statements about computer programming, software development, etc. since you don’t seem to have a background in the field, and aren’t working from any specific Tesla statements. For example, you seem to think that neural networks don’t use software, which I assure you they do. At the bottom of the pyramid there’s a processor (actually many in this case) executing instructions, i.e. “software”. Granted the NN architecture is very different from classical procedural programming, but it’s still software. | 24/04/2019

@Dramsey - makes sense, although most "promises" were never promises but goals, but agreed it's hard waiting for the next feature or improvement. Not disputing some promised deadlines were missed too. Clearly EAP and FSD is a lot harder than anyone expected. Seems they are still way ahead of anyone else though. Ah, the fun of being on the bleeding edge :)

Well worth watching the AP video from Monday if you're at all technically inclined. Never seen a tech company provide so many inside details of what they are doing. (skips over the pre-video, but you can start from the beginning if interested). Tech stuff starts around 1:10 mark.

Dramsey | 24/04/2019

I watched the entire AP video. As an aside, Tesla should do something next time about mounting their presenter-visible screen on the floor, tilted up. I thought the chip designer guy had autism or something since he was always looking down rather than out at the audience. It wasn't until later in the video that I noticed where the screen was mounted.

I admit snickered at the part where Elon earnestly said that ALL CURRENT TESLA CARS HAVE ALL THE HARDWARE NEEDED FOR FULL SELF DRIVING. Hm. That's what he said when I bought my car. I know, I know, I'm getting a free computer upgrade. But still.

(Wonder what, if any, capabilities I'll miss out on due to my vehicle's monochrome cameras?)

Anyhow. Back to patiently waiting for 2019.12.x for my car. I WANT NON-PROMPTED LANE CHANGES DAMMIT.

dougk71 | 24/04/2019

@Dramsey, to dougk71
I think you need to be more careful when you make definitive statements about computer programming, software development, etc. since you don’t seem to have a background in the field,
I had 300 programmers reporting to me and my software company went public in 1983. Since them I have worked with VHDL and Verilog mostly for specific projects that I pursued given that I retired very early. I am a Mathematician and yes I understand Markoff chains linear algebra and matrix dot products.
Anyway it is not what I said or about me but what Tesla's software and hardware engineers said in their autonomy presentation. Now you may be skeptical and that's your choice but my money ( yes I own Tesla shares) that Tesla software experts no more than either of us. Tesla is moving away from NVIDIA to gain better performance for FSD. I have paid for FSD and welcome an upgrade to a chip that is 21 times faster than the current NVIDIA chip set in HW2.5
If all you can say about the guy that developed a Neural Net chip 21 times faster then a NVDIA HW2.5 chip is that the looks autistic to you tells me that you understand little about what an achievement that is.

bill | 24/04/2019

@Dramsey "But remember that Elon said for years that HW2 was all that was needed for full autonomy (it was even on their ordering web page, something along the lines of "Every Tesla comes with all the hardware necessary for full self driving capability", before changing his mind and announcing HW3."

I am pretty sure he said at the time that the CPU may need to be upgraded but it was designed for that.

Dramsey | 25/04/2019


Well, to be clear, I wasn't criticizing the content of Tesla's HW and SW presentation, I was criticizing the manner of the presentation. Compare it to a typical TED talk-- or, to see what a REAL presentation looks like, one of Steve Jobs' classics).

When you're trying to convince a room full of big investors (as opposed to a room full of engineers) in your company that what you've got coming down the pike is amazing, presentation counts for a lot. Deep dives into chip and software architecture are all well and good, but these are people you want to give you LOTS OF MONEY, and you need to make them excited about your product. Tesla's presentation was not exciting.

And although I believe in Tesla to the degree that I've bought two of their very expensive cars, my experience with their promises and pronouncements since 2013 when I bought my first Model S has led me to be very skeptical of these "forward looking" statements.

Glad to hear that you've a lot of experience with programmers and engineering. I spent over 30 years as a professional programmer myself.

But I'd guess that, like me, you know virtually nothing about modern NN design and programming. This makes it difficult for us to really evaluate the claims they're making. We have to take it on faith that their new chip is a zillion times (or whatever) faster than NVidia's Drive PX2. Furthermore they made this claim without offering specific benchmark figures, other than very general ones like "processes X more frames per second."

I was a little confused when you seemed to say in an earlier post that NN as implemented on Tesla's new computer "wasn't software", which is why I commented on it. Forgive me if I misunderstood your point.

Dramsey | 25/04/2019


When I bought my current car in late 2016 and paid up front for FSD, the order page was very clear that my car had all the hardware it needed for FSD as delivered. Elon started hedging his bets by mentioning a possible computer upgrade about 6 months later, if memory serves.

bill | 25/04/2019

Bottom line is they sold us FSD so they owe us FSD,

dougk71 | 26/04/2019

Individual stock analysts come and go. The organizations they work for often lack the business expertise to stay in business...Drexel Lambert and Rothschild to name a couple
. It is not that other Automotive stocks had a great 2019 Q1 either.
I do think Tesla needs to advertise its cars to expand its acceptance and cement the brand. If all you have is bad publicity and nothing to push back with you can get wrongly branded. Even on pro Tesla You tube channels ICE cars are advertised

sreeni848 | 26/04/2019

Does anyone know when Tesla starts replacing the old CPU's (HW 2.5) to new ones (HW 3.0)?

Dramsey | 26/04/2019


"Bottom line is they sold us FSD so they owe us FSD,"

True, and Elon has said, many times, that HW2 people who pre-bought FSD will get a free upgrade to the new computer. However, Elon's said a lot of things, and historically receiving promised Tesla stuff (referral awards, etc.) has not been...quick.

Since Tesla's finances seem a little fraught just now it wouldn't surprise me if all new HW3 computers went to new cars for quite a while before we receive ours. Of course it's all a moot point until FSD features that require HW3 are available. As I noted earlier, I'm still waiting for 2019.12...

EVRider | 26/04/2019

@sreeni848: Tesla won't start upgrading cars with HW2/2.5 to HW3 until they introduce FSD features that require the upgrade. Elon indicated that might be a few months away.

Bill_75D | 26/04/2019

I worry that once FSD features become available, my 2016 prepaid FSD car will then start to wait for the upgrade parts coming "soon".

Why not give me the upgrade now so I'm set when they begin to release features? After all, I paid for it 2.5 years ago.

EVRider | 26/04/2019

@Bill_75D: Why should Tesla spend money now for an upgrade to your car that won't do anything for you until later? Especially when "later" is Elon time? Tesla needs to focus on being profitable. It doesn't matter whether your car is a 2016 or early 2019 -- all of us who bought FSD are in the same boat.

Bill_75D | 26/04/2019

You're right. Besides, HW 4.0 might come out by then and I'll need another upgrade.

Haggy | 02/05/2019

I agree that 5% or less sounds reasonable for the take rate, at least prior to the current offering. Now that there's no EAP, there might be a higher percentage taking it up front since there's already a big difference but those people won't need the board upgrade.

Most people with the old processor got it when EAP was being sold, and I wouldn't be surprised if the number of upgrades needed is under 20,000. I wouldn't be surprised if it's as low as 10,000.

gregmiller1817 | 11/05/2019

2014 MS P85+. Still hanging in there for FSD. That's why I figure when they finalize hardware and software, I'll upgrade/trade-in to FSD. In the meantime, Love my MS. What a machine! I'm an old hotrodder and would love to update wiring harness and bolt-on's to bring mine up to the present. Can you imagine a car updatable to current year. Where's the legacy upgrade kit? I'll buy it! The P85+ is sitting right next to my 1930 Chrysler model 70. What a contrast.

inconel | 11/05/2019

I'd say 1% chance of upgrade via harness kit, 99% via trade-in :)
FSD option might become more expensive when they have more features though.

bp | 11/05/2019

Since the FSD processor is manufactured in China and just got hit with the 25% tariff, wouldn't be surprised to see Tesla defer the FSD upgrades while they wait for the tariffs to be reduced or eliminated.

Boonedocks | 11/05/2019

It is not made in China. It is however made in Texas

EVRider | 11/05/2019

Regardless of where the manufacturing takes place, Elon said they were using Samsung, which is Korean, not Chinese.

Tropopause | 11/05/2019

Tesla applied for, and was denied, an exemption to the Chinese tariff for some chips that are supposedly used on the new HW3 board. Something on the new HW3 board is made in China and, according to Tesla's petition, it is coined "the brain" of FSD.

Tesla also mentioned Samsung in Austin, TX as another supplier of HW3 components. | 11/05/2019

The HW3 has hundreds of components, most are only a few cents, but a few are quite expensive. I'm sure a few components come from China, but whether these are the $100 parts or $0.10 parts is not known. Some of the most expensive parts do appear to come from outside China. In the past, the entire PCB assembly for major items like the MCU has been made in Fremont, CA. Don't know if that's still true for the HW3 board, but I suspect so.

bp | 12/05/2019

The FSD chip is made at Samsung's plant in Austin. It appears the circuit board on which the chip is placed is made in China - because only that company was able to meet Tesla's aggressive schedule.

"Tesla was unable to find a [U.S.] manufacturer with the requisite expertise to produce the Autopilot ECU 3.0 with the required specifications, at the volume requested and under the timelines necessary for Tesla's continued growth." | 12/05/2019

@bp - Where did you find that quote? No reason to suspect it isn't true, but I'd expect the HW3 PCB to have a "Made in China" mark, which I couldn't find. There is a TTM logo (TTM Technologies), who make PCBs and assemblies. They are headquartered in California (although they could make product in China).