TSLA Market Action: 2018 Investor Roundtable

[MissAutobahn]

@Fact Checking - you seem to post 24/7/365 and have an in-depth knowledge on pretty much everything. Are you a gestalt entity?

As long as I have the feeling @Fact Checking is also running on Tesla AutoPilot, I feel safe that his thoughts and facts are clear and save even without sleeping

 

[NicoV]

Tesla will also be able to use multiple AI chips in the future: for example two discrete AI chips on a single board handling 4-4 cameras.

(Note how the processing workload of 8 cameras can be split into 2, 4 and 8 parts, allowing a lot of future hardware parallelism - I don't think the number of cameras was an accidental design choice...)

They could also alternate between two chips on each frame, at a cost of 1 frame of delay if they get over the capacity of a single chip solution. I would prefer a new chip with twice the processing power, it could leave the rest of the system without changes.

I think an important but underappreciated advantage of the Tesla chip is the power advantage. GPU’s may use hundreds of Watts, while the Tesla chip may use a lot less. In an robotaxi situation where the chip runs 24/7 full tilt, the cost advantage in power may be much bigger than the cost of the GPU itself.

 

[Joe F]

Already debunked as false on CNBC.

Looking for a link, but none found. It was a blurb on-screen live, without further info. Might have been a mistake as the only info on-line is that the deal was pulled.

Regardless, as had been stated, there's no connection to Tesla.

 

[Tslynk67]

Use your work’s laptop then. Me too, I can’t believe I just read an SA article.

That's it, you have the mark of the demon now! You're accursed! You must pass the parchment to another before midnight!

 

[BenPrice]

Correct me if I’m wrong, but Tesla now has started the 45 day clock to install new chairman?

Wonder if that will occur before or after earnings...

 

[MP3Mike]

Right now they process everything, all frames from all 8 cameras with a single discrete GPU I believe, on an Nvidia GP102 based board.

Actually for v9 it appears that they had to tap into the integrated GPU as well. From Elon:

To be clear, actual NN improvement is significantly overestimated in this article. V9.0 vs V8.1 is more like a ~400% increase in useful ops/sec due to enabling integrated GPU & better use of discrete GPU.

So maybe they are running two cameras off the integrated GPU and six off of the discrete GPU?

 

[X Fan]

So popular mechanics has a new article titled "In defence of Elon Musk". I spent around half an hour reading it and my perspective has changed on the significance of some of the more erratic parts of Elons last few months. Even though I always supported him and understood he was under immense pressure from many sides, I was somewhat concerned about what was going on with the pedo tweets and the lack of a formalised approach to "funding secured ". Honestly, the last few months now feels like a footnote in an incredible story that is about to unfold. Nothing more. Please read if you have the time.

In Defense of Elon Musk

Great article, thanks for sharing!

 

[humbaba]

The action is now at $275... A nice, healthy ~6% increase in value. Despite a strong start, trade volume tapered off leaving today with less than average (yeah, its not closed yet, but its trailing average by 2.4M trades).

Barring some bizarre late afternoon event, this looks like a solid day for $TSLA. If it continues then the stock will soon return to previous levels.

Of course, this is $TSLA, so... yay for volatility! But, as with the rest of the stock market, it appears to be in recovery.

 

[TradingInvest]

"Fact Checking said: ↑
(Note how the processing workload of 8 cameras can be split into 2, 4 and 8 parts, allowing a lot of future hardware parallelism - I don't think the number of cameras was an accidental design choice...)

Sorry but can you explain this part for me? Thanks.

While I normally agree with "Fact Checking", not on this point about future hardware parallelism. Tesla has 8 cameras to cover 360 degrees around the car. 7 won't do the job, 9 seems unnecessary.

 

[Tslynk67]

It’ll be almost certainly stress weight, along with a general body, hormone, sleep etc... imbalance.

Now that that’s resolved - good morning opening peak, welcome mmd.

Indeed, cortisone must be all over the place.

@Fact Checking - you need to get more sleep, dude, get your circadian sorted out. Don't think of it as lost time, if you fix your sleep then you'll be even more productive in the waking hours you have, and even sharper.

 

[NicoV]

I believe the 'cost' there refers to the cost to customers, i.e. no price hike necessary due to HW3.

It is extremely unlikely for an entirely new piece of hardware to have the same cost as the old hardware. The more probable explanation is that the new hardware is cheaper, so there's no cost/price increase to customers.

But that's just speculation, maybe there's some cost I'm overlooking:

• For example NVidia might have given Tesla a really sweet deal for the GP102 chips (Pascal micro-architecture based), as a bait-and-switch for the much more expensive Xavier based chips they are offering currently.
• Also, I estimated the direct marginal manufacturing costs of the new MCU board, while Elon might have included the very significant R&D costs - it will be some time until the AI chip recoups the money invested into developing it.
• Or the Tesla AI chip might be using some bleeding edge fab process that is significantly more expensive than the couple of dollars I estimated.

Or the Tesla AI chip has a monster size and/or low yields. Given that this thing has to run in a harsh automotive environment and requires a combination of high performance and low power, it may be an SOI chip, which is also more expensive.

 

[Tslynk67]

It’s all about the batteries

Fire Away!

Indeed; batteries, fire, away! Thank you.

 

[humbaba]

After publishing a fair and balanced article on Tesla, Claudia Assis is back to the usual. Rather than report on the stock recovery lets find something negative to talk about...

Here’s how much Tesla lost in market cap now that Musk’s SEC settlement is approved

(not that I read it, the title was enough for me)

 

[homer214]

Removed "like" rating.
Added "love" rating.

I smiled while reading every word. Love the connection to Twain and others - warts and all - who have made a difference. It is and will continue to be a helluva ride.

 

[wipster]

Ha ha, Keef is the first post. What a tosser!

Edit: on further investigation, I think it's Keef's page. What a tosser!

Boy, they sure get a lot of posts... NOT! Ha!

 

[Thekiwi]

Meh - I think everyone assumes it is a done deal and I don’t think confirmation from judge will be market moving.

I am very happy to be completely 100% wrong on this one!

 

[mickificki]

Yeah, so visual input processing is the most computing intense part of full self-driving. Tesla has 8 cameras, and if you want to process each at 100 fps (one frame every 10 milliseconds), at the native HD resolution of the cameras, that's a lot of processing.

Right now they process everything, all frames from all 8 cameras with a single discrete GPU I believe, on an Nvidia GP102 based board.

But now that they have their own discrete NN chip, the Tesla AI chip, in future iterations (HW4, HW5) they could use the following computer topology within the board, with very little additional cost (the AI chips probably cost only a few dollars to make each - most of the cost is in making the board):

   [AI Chip #1]           [AI Chip #2]
               \         /
                [GPU RAM]
               /         \
   [AI Chip #3]           [AI Chip #4]

I.e. four chips and shared RAM of say 16 GB high-speed GPU RAM with multiple access channels so that all CPUs can use the RAM all the time without slowing down each other.

(There's also the question of whether the Tesla AI chip uses separate RAM modules - a possible alternate design would be for the RAM to be integrated into the AI chip itself, as a sort of very fast transistor based SRAM. This would have a number of other advantages as well, such as close proximity of NN 'weight' data with the functional units representing 'neuron' nodes.)

But assuming that RAM is separate from the chip, the above board layout is a possible topology, where Chip 1 would handle cameras 1-2, Chip 2 would handle cameras 3-4, etc. While not all cameras have the same pixel count, the processing overhead is still similar and scales with the complexity of their neural networks.

Note that this way the total computing throughput of the system can be increased by a factor of 2x, 4x and 8x with very little additional cost other than a higher power envelope.

I'm reasonably sure HW3 is going to feature one AI chip (they want to keep it simple initially, and it appears the chip is plenty fast already) - if it features two chips it will be for redundancy and fail-over perhaps, not to increase performance.

All of this is speculation though - I'm sure we'll hear more about the details once the HW3 release gets closer ...

Thank you! really appreciate the work you're doing here!

 

[davecolene0606]

9 (cameras) takes us inside the car on the 3

Fire Away

 

[capster]

VW CEO says German carmakers have only 50% chance of staying ahead

They realised in which position they are, now confirming it and face it.
I think that their chances are lower.
What investors in legacy carmakers , MSM, shorters and other infrastructure think now?

50% is over optimistic: "Objects in the mirror are closer than they appear."

 

[Fact Checking]

Actually for v9 it appears that they had to tap into the integrated GPU as well. From Elon:

"To be clear, actual NN improvement is significantly overestimated in this article. V9.0 vs V8.1 is more like a ~400% increase in useful ops/sec due to enabling integrated GPU & better use of discrete GPU."​

So maybe they are running two cameras off the integrated GPU and six off of the discrete GPU...

Edit: I think you are right!

This is the HW 2.5 layout of the Nvidia board (AutoPilot ECU):

The "discrete GPU" Elon refers to is the "Pascal chip": 12 TFLOPs monster with 3,328 CUDA cores.

Note the four black squares above the Pascal chip: those are the GPU RAM chips, 2 GB each, 8 GB total. They are close to the Pascal chip for maximum performance.

Edit/Correction:

According to @MP2Mike the Parker chips include a 256-core Pascal integrated GPU as well:

The Parker chips also include 2+4 superscalar ARM cores, for execution.

But there's also a second board in the Tesla HW 2.5 computer module, an Intel board (the MCU), reportedly with an Gordon Peak and Apollo Lake SoC's: I have not found any tear-down images - the board's more interesting side is sealed inside the computing module.

Intel Apollo Lake Atom CPUs do come with an integrated GPU: but its performance is not comparable to that of the Pascal chip - it's maybe 2-5% of its performance. So I doubt it's used to process video data or do any significant AutoPilot work - it might be used for smoother infotainment graphics though.

I think all of AutoPilot runs on the Nvidia board: the neural nets execute in the discrete Pascal GPU (GP106 based), and the two Parker chip integrated GPUs, processing input from all 8 cameras, the sonars and the radar in that single chip, and the actual high level vehicle control logic (which uses the continuous stream of NN output) executes on the ARM cores of the two Parker chips.

Much (~86%) of the processing power is in the discrete Pascal GPU, but there's another 7%-7% in the Parker chips as well, in form of smaller integrated GPUs.

The Intel board is used for visualization and all the apps and infotainment processing. It has an integrated GPU for smooth graphics: a HD Graphics 500 or 505 integrated GPU.