Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable

This site may earn commission on affiliate links.
Agree on I/O compatibility but.... it is not the same board as the NVDA board and it needs testing and validation particularly on such complex software that is hard to even understand.

The code is not hard to understand though. It is purely a series of multiply, adds, max clipping and such run on a huge array of numbers. The memory handling is likely different, but block copies are not complex either.

What is complicated is what the NN is doing internally, but that doesn't change based on HW.

They can run the two versions of HW through the complete test data set. If the outputs match bit for bit over multiple runs, that is high confidence of equivalence. That is what FC meant by
HW3 is very likely "100% binary input/output compatible" with HW 2.5 neural networks, IMO.
Not that the HW I/O matches, but rather that the same input data stream with the same NN produces the same output data stream.
 
Just don't be upset if it turns out not to be true. I doubt very much any rep is authorized to talk with a customer about HW3 availability.

It's easy to get a customer representative to say what you want to hear. Especially an enthusiastic one who is wanting to please their customer with good news.

And usually Tesla customer representatives are kept in the dark. It is unlikely that they know anything about when HW 3 comes.
 
MODERATOR:

Please place ALL posts regarding
Tesla cargo ships here: Tesla Cargo Ship (Thank you, Buckminster).

We are having trouble porting over to there the ones posted here this morning. IF one of those was yours, please go copy-paste your own post over there, after which Moderators will delete the original ones in this thread.

Can I go back to sleep now? Tough night......

I think I've copied all of mine over. Hope you're okay! :(
 
Agree on I/O compatibility but.... it is not the same board as the NVDA board and it needs testing and validation particularly on such complex software that is hard to even understand.

Note that even with HW 2.5 Tesla was already making their own board, it's a custom Tesla board, not an Nvidia board.

The HW 2.5 board contains about 4 Nvidia chips, one of which is an automotive GPU chip, used by the general purpose chips on the board.

In HW3 I expect Tesla to be using much of the same board layout, the same (well, similar) ARM based general purpose chips, but the Nvidia GPU chip replaced by Tesla's own ASIC NN acceleration chip. (And probably there's more and faster DRAM chips as well.)

I.e. the board is plug-in compatible not just with the rest of the system, but is also binary compatible on the OS and application level, to almost every degree but the NN acceleration libraries.

So while it's obviously a big hardware upgrade, they very likely kept the software compatibility risks as low as possible. This was very likely extensively tested.
 
Hi all,
my 1st post, hoping it to be potentially useful. Today's the 2nd time I see something about HW3 regarding European M3 deliveries.

In short: My Model 3 LR March delivery to Europe/Finland will have HW3.

In long, attempting to establish credibility level: About two weeks ago I called Tesla customer service Finland (redirected to Norway due to queue). Asking if my M3 will have HW3 sent the customer supp rep to discuss with someone what's the deal. Coming back in a couple of minutes confirming just what I wanted to hear. She did sound sincere and very enthusiastic through the whole discussion.

Other points that were not clear on my Tesla account: the all season weather tires will be Nokia R3 (exactly what I wanted); FSD can not be added to my order (oh well, later then).

And.. many thanks to all contributors of the thread! It's been a valuable tool with my TSLA investments.

It's possible your Model 3 will have HW3 but I wouldn't put too much credence on what Tesla customer support says. They really don't have any inside info on company plans.
 
Yep, and I got one or two of those calls from Las Vegas Tesla sales people, telling me exactly that, about the $7500 Federal Tax Credit ending soon (I am so sick of hearing about it frankly) and I ought to get my Model 3 and why haven't you yet I mean you've had your reservation since, let's see--"March 31, 2016"--we both said in unison. I told him I simply don't care about the tax credit, and I am sitting this Model 3 craze out until the State of New Mexico gets a Tesla store and service center, even just one, so that we don't have a two-Tesla household with the closest service center 400 miles away, the risk for us just feels too high, considering how much service my S has required over the years. So the sales guy said "well there's no point in us holding on to your $1000 reservation anymore because it's not needed" and I said "okay well how fast can I get it back" and he gave me all these different options including a gift card (!? hell no) and I said I look I just want a check and he said oh well that is the slowest method, like up to 60 days' wait. So I said well hold on to the reservation a little longer as we still wanna buy the car but we're waiting, like I said, and also, we wanna see Tesla pump out 300,000 or 400,000 3's out of the factory before we get one, just to make sure Tesla's worked out a lot of the glitches, again given the number of issues, most admittedly minor, but all requiring service (including multiple drive units and replacement of the entire DC charging system). The sales guy, his talking points exhausted, surrendered, and we said bye and that was that. I still eagerly await getting a 3 but we're sitll in waiting mode. I wonder how many other people Out There are in a similar situation--absolutely positively getting a 3 eventually, especially if they're already an S owner, but taking their sweet old time (which is driving Tesla crazy).

I’m in similar situation. After putting deposit on 3 I leased S(60D HW1). And I honestly love it so much, I don’t see myself “downgrading” to 3 when lease expires in August. Wife said categorically NO to second Tesla (for now, I’m still working on her;)), so not sure what I’ll do.

Current plan is to buy 100D HW3 when this lease expires and roll the 3 reservation into Y with hope of convincing wife for 2nd Tesla by then.

Not sure if they will allow conversion of reservation like that.

Edit. P.S. And yes I got that call too.
 
Watching the RO-RO ships travel around the world, something struck me.

I've written a couple times (here and elsewhere) about how while it's not economically practical (with current tech) to make large electric-powered cargo ships that travel nonstop (don't care to redo the calculations yet again, but feel free to do them yourselves), it is economically practical to use them with floating "gigachargers" (deep sea wind, floating solar, inside a breakwater - ideally with the breakwater being a wave-power generator). These would transfer - for ships the size of a Maersk Triple-E - about a gigawatt hour per 80% charge, about every day or so.

Something occurred to me, though - and the situation could actually be a lot more than merely "economically practical" - rather, a major economic advantage.

Speed is a key part of the economics of shipping. For one, the faster you deliver your cargo, the more trips you can take. For another, the faster you deliver cargo, the more you get paid for that delivery (the reason why people do air shipping even though it's insanely expensive compared to shipping at sea). Double the speed and you might quadruple your revenue, for a given capital investment.

So why don't ships just go faster? Energy consumption, of course (operations, not capital, costs). The faster you travel, the more energy your ship has to burn to do so. Ships today don't want to have to pay more for fuel, so their cruising speeds are limited (the Glovis fleet usually cruises at about 20mph/30kph, for example).

Now, this might seem even worse for electric shipping. After all, big batteries are expensive, and the more power you burn, the larger the battery you need to have in order to charge at a given interval. But what happens if we reduce that interval significantly?

The rate at which you can charge a battery pack is irrespective of the size of the battery pack; for a given cell and cooling design, a 1kWh pack takes the same amount of time to charge as a 1GWh pack. A ship can do the same 30 minute 0-80% that a car or truck can, so long as the charger are sized to do so (just through a *much* fatter, crane-hoisted cable!). You certainly have more overhead - sailing a ship into a breakwater, docking alongside a charger tower, and connecting a liquid-cooled cable wider than your thigh, is not a 1-minute job like parking your car at a Supercharger and plugging in. But assuming that overhead can be kept "reasonable", there's nothing to stop you from charging far more often than once per day.

(Note that electric propulsion makes things like azimuth-mount thrusters ("azipods"), which allow ships to sail sideways and tightly control their position, more practical)

Let's say that instead of sailing for 23 hours and docking / charging for 1 hour, you sail for 5 hours then dock/charge for 1 hour. Now you're charging 4 times as much energy per day, for 87% as much sailing time. Burning four times the power lets you roughly double your travel speed - ~40mph/60kph. Meaning you can depreciate your capital costs across far more trips, and get paid more per trip for the faster delivery speed.

The only downside is that you burn twice as much power per trip. From an environmental standpoint, it's really a nothing issue: it's the power of the wind and/or sun, and most of the world's oceans are "deserts" - vast expanses with relatively little life, due to the lack of the sort of nutrient upwellings that you get near the coasts:

Seawifs_global_biosphere.jpg


In the above map, dark red zones have 1000 times more photosynthesis as dark blue zones, 150 times more than cyan zones, and 50 times more than green zones. It's mineral-limited, not sun-limited; if you block some sun in one location, it just leaves the minerals for the next bit over. On the other hand, sea life tends to flourish around manmade floating structures, akin to how it does around reefs.

Historically, ships have been getting a great rate on fuel costs, as they've been burning high-sulfur bunker fuel. Those days come to an end at the end of this year - the standards on bunker fuel have been raised to the point that it's now basically diesel, and in direct competition with diesel to boot. Ships can still use low grade fuel, but only if they put in (expensive) scrubbing systems on their ships that may cost more than just switching fuels. You're looking at at least "$2/gal" equivalent (prices are usually measured in $/MT), and more if oil prices rise from their (currently low) pricing regime, or further emissions restrictions (or carbon taxes) increase costs further. Let's say a long-term average of $2,50/gal - and that may well prove incredibly optimistic in the long run.

Ship engines are efficient - about 50%. Now, EV motors would also be unusually efficient in such situations, as they'd be large motors tuned for cruising speeds, and the charging process would also benefit from operation at scale. Let's say 87% round-trip efficiency. The fuel-powered ship gets propulsive energy for 27MJ/$. So if we're doubling the propulsive energy requirements, in order to match the price, electricity (at industrial rates, not home rates) needs to be generated at 54MJ/$ - aka, $0,067/kWh. Remember that it doesn't actually need to match bunker fuel costs, as you're shipping at nearly double the rate, drastically slashing your depreciation per trip while drastically increasing your income per trip.

That said, it would be awesome if electricity costs could beat fuel costs even when moving at double the speed. Is $0,067/kWh achievable? Well... "probably"?
  • Floating solar plants have so far mainly been built in freshwater, but if you have an effective breakwater, then it just comes down to an issue of material compatibility. Prices are similar to that of land-based PV - for example, Three Gorges Group is making a 150MW floating solar plant for a construction cost of $151M, or $1/W. That's just a few cents per kWh generated. The fixtures are more expensive, but installation is simpler and cheaper, on cheap/free "land". Since floating solar is a newer technology, it also has more room for price improvement.
  • Deep sea wind is not yet there in pricing; it's currently significantly more expensive than land-based and shallow-water wind. That said, it's also highly immature, and has a lot of room for improvement (and all oceanic wind has the advantage of being basically unlimited in tower height, with hardware shipped cheaply to its destination). Additionally, one of the major costs of deep-sea wind is transmission back to the shore, which is not applicable here.
  • Wave power is currently expensive, but regardless, not much is needed - only enough to make a breakwater.
Floating solar, at present, looks like the most realistic option for the bulk generation, with deep sea wind only as a supplement (turbine towers could double as platforms for storing charging hardware and/or docking ports)

Can chargers (and battery banks) be built at scale, using adjacent-generated solar at current solar pricing, and sell power for $0,067/kWh? That's harder to say - but this is exactly Tesla's plan for megachargers for Semi - and their announced pricing is $0,07/kWh (combining the low cost of solar generation with the battery banks it needs to be a reliable power source (direct DC/DC conversion, no grid costs) - batteries which simultaneously enable high charging speeds using said same DC/DC converters). A gigacharger would gain even larger economies of scale.

So... "probably". But the key aspect is: you can earn drastically more revenue from your ship if you run it on electricity, by sailing faster - since your fuel is cheap, clean, and it's much cheaper to add more electric powertrain power than diesel power.
Just responding from the peanut gallery: ABSOLUTELY LOVE that proposal. Truly mind-boggling, in a good way. I only wish I could double-tap that heart symbol!
 
HW3 is very likely "100% binary input/output compatible" with HW 2.5 neural networks, IMO.

This makes it much easier to deploy: HW3 produces the exact same output when running the same networks as HW2.5, if fed the same input data (video frames).

Installing HW3 right now avoids the expense of having to upgrade it in the near future, and probably also lowers costs, as Tesla doesn't have to buy Nvidia GPU chips anymore.

HW3 will truly shine once the much larger, more sophisticated FSD networks are released, later this year - and that can be done via an OTA upgrade.

While I agree that HW3 will have the same input/output, the complexity lies in the fact that a NN running on the TRIP hardware will be a completely different binary than what runs on the nVidia GPU today. Since this is a brand new architecture, I would imagine that Tesla themselves had to build a compiler to translate the NN SW from their development framework into the "compiled/minimized" binary that is executed on the TRIP hardware. So the key question is how far along Telsla is in validating existing Autopilot features on HW3. I bought my Model S with AP 2.0 in 2017 and it wasn't until about a year later that Tesla shipped V9, finally delivering software capable of exploiting the AP 2.0 hardware. I'm sure Tesla will not want to repeat the HW2 experience with HW3.
 
While I agree that HW3 will have the same input/output, the complexity lies in the fact that a NN running on the TRIP hardware will be a completely different binary than what runs on the nVidia GPU today. Since this is a brand new architecture, I would imagine that Tesla themselves had to build a compiler to translate the NN SW from their development framework into the "compiled/minimized" binary that is executed on the TRIP hardware. So the key question is how far along Telsla is in validating existing Autopilot features on HW3. I bought my Model S with AP 2.0 in 2017 and it wasn't until about a year later that Tesla shipped V9, finally delivering software capable of exploiting the AP 2.0 hardware. I'm sure Tesla will not want to repeat the HW2 experience with HW3.
They seem to have an aggressive in-house testing program for HW3 and autonomy (plenty of information on that late last year). I think they are completely on track to a wide release of a well tested HW3. Although I was expecting a full horn-blowing release announcement for HW3, they might be doing a stealth rollout of the hardware and saving the horn-blowing for the next software release.

(Not that I expect that to bring autonomy, but that is a longer subject for a different thread.)
 
  • Like
Reactions: Fact Checking
I’m in similar situation. After putting deposit on 3 I leased S(60D HW1). And I honestly love it so much, I don’t see myself “downgrading” to 3 when lease expires in August. Wife said categorically NO to second Tesla (for now, I’m still working on her;)), so not sure what I’ll do.

Slightly different position: I purchased an S instead of the 3 last year, partly because I didn’t think my wife would put up with a “downgrade” in road trip car size from our Subaru Outback hatchback. She really still prefers to drive the Jeep Rubicon, and doesn’t want 2 EVs. SO my current strategy is a deposit on the Rivian R1T, and I have gotten her jazzed about how much better an adventure vehicle that will be than the Jeep, and yet still a good road machine. Still likely almost 2x cost of a 3, but gets me to an EV-only garage, and I love the specs on that 4-motor R1T. Stay tuned.
 
While I agree that HW3 will have the same input/output, the complexity lies in the fact that a NN running on the TRIP hardware will be a completely different binary than what runs on the nVidia GPU today. Since this is a brand new architecture, I would imagine that Tesla themselves had to build a compiler to translate the NN SW from their development framework into the "compiled/minimized" binary that is executed on the TRIP hardware.

Since they actually designed the hardware, I don't think they built a 'translation compiler' - instead they extended their NN libraries with HW3 support, and tested it extensively that HW 2.5 NNs are upwards compatible on HW 3.0.

Their NNs are effectively big files generated by their NN training clusters and which are part of the firmware upload and which are never modified by the car software itself, with the neural network topology and weight data in these files.

These are relatively simple data structures which are uploaded to whatever NN accelerator hardware is available:
  • On the NVidia GPU they probably had special-purpose pixel/texture shaders (or straight CUDA code) that processed neural network data: glorified matrix multiplication with saturation clipping in essence, plus some sort of mailboxing architecture to send frames to the GPU and receive processed output data in exchange.
  • On the TRIP (HW3) hardware they probably have a native compiler for that architecture which generates native binaries that do the inference processing and the mailboxing.
Their NN libraries are probably hardware-agnostic: same NN files result in same input/output results when using the NN library, regardless of whether it's running on HW2, HW2.5 or HW3.

This kind of design decouples their NN acceleration hardware from their software. This is possible only because inference processing is mathematically and computing wise very simple, compared to the complexity of the rest of the system.
 
So my current strategy is a deposit on the Rivian R1T, and I have gotten her jazzed about how much better an adventure vehicle that will be than the Jeep, and yet still a good road machine. Still likely almost 2x cost of a 3, but gets me to an EV-only garage, and I love the specs on that 4-motor R1T. Stay tuned.

R1T did look VERY interesting. Not sure I’m in market for pickup, but it had a lot of nice design touches. I’m very curious how competitive Tesla truck will be with that thing.
 
  • Like
Reactions: VValleyEV
While I agree that HW3 will have the same input/output, the complexity lies in the fact that a NN running on the TRIP hardware will be a completely different binary than what runs on the nVidia GPU today. Since this is a brand new architecture, I would imagine that Tesla themselves had to build a compiler to translate the NN SW from their development framework into the "compiled/minimized" binary that is executed on the TRIP hardware. So the key question is how far along Telsla is in validating existing Autopilot features on HW3. I bought my Model S with AP 2.0 in 2017 and it wasn't until about a year later that Tesla shipped V9, finally delivering software capable of exploiting the AP 2.0 hardware. I'm sure Tesla will not want to repeat the HW2 experience with HW3.

I think you are talking about three different things:
1. The complier/ interface to run a NN on the new HW. This would have been completed as part of the chip development/ validation.

2. AP1 to AP2
AP1 to AP2 was a sudden hardware switch with no preexisting NN.

3. The NN itself
The NN currently running v9 on HW2.x can be run on HW3, so there would not be feature changes in the transition.

The next step change will be FSD on HW3.
 
Am I the only one freaking out about the fact that Tesla has had to lower the price twice since the start of the year? I really thought, at least in Q1, that Tesla would be able to sell out 6000 to 7000 per week at $46,000 and up in the face of the large European and Chinese backlogs.

I think it’s great that Tesla can reduce the price because of process improvements while the Federal tax credit is still in affect (albeit 50%). They certainly are still selling well.

All EV manufacturers should lower their price when they can if they can still be profitable. The tax credit isn’t meant for car companies to keep that as extra profit to make EVs. I applaud Tesla for doing this.
 
Interesting insight about e-tron, Taycan versus Tesla battery cooling systems and the implications.

Either Audi & Porsche are just smarter than Elon which is a possibility but hard to believe if we talk Batteries or they did make some marketing statements and users won't have long fun with it...

Audi e-tron Battery TMS: How Does It Stack Up Against Tesla Model 3?
 
Although I was expecting a full horn-blowing release announcement for HW3, they might be doing a stealth rollout of the hardware and saving the horn-blowing for the next software release.

Right now they obviously don't want to Osborne Effect themselves, that's why they are mum about HW3, even if it's already in EU/China models.

Even in early April they might not want to go full horn-blowing - only once the first FSD features like stop signs, traffic lights and left/right turns are released. Then I'd expect full horn blowing, coast-to-coast FSD trip and the whole nine yards.

Because that's when HW3 will start generating more sales and more revenue for real.
 
Last edited:
Interesting insight about e-tron, Taycan versus Tesla battery cooling systems and the implications.

Either Audi & Porsche are just smarter than Elon which is a possibility but hard to believe if we talk Batteries or they did make some marketing statements and users won't have long fun with it...

Audi e-tron Battery TMS: How Does It Stack Up Against Tesla Model 3?

You know, I'm tired of dancing around the "buts" and "maybes" and "what ifs". Let's just call it as it is: Porsche and Audi are willingly frying their cells, because most people won't fast charge that often and they can afford to eat the warranty costs on those who do.

There is no other explanation that meets Occam's Razor. There is no magic cell that somehow only they have that nobody else has heard of, that manages to be simultaneously energy dense and power dense. Their cooling design is primitive. They're not "locking out" half of the pack to make the C-rate appear faster than it is.

Let's call a spade a spade: they're frying their cells.