US$15.000 NVIDIA DRIVE PX 2 supercomputer in every Tesla?

[Frogwheels]

Correct me if I'm wrong, but from what I've been picking up, the new autonomous capabilities in every new Tesla vehicle are powered by an NVIDIA Drive PX 2 processing unit.

Each unit is liquid cooled supercomputer and as powerful as "150 Macbook Pros" (quote from NVIDIA CEO at CES 2016).

BUT the list price of this unit is US$15.000!

NVIDIA CEO Jen-Hsun Huang seems to be an admirer of Elon, so let's say he grants Tesla a discount and gives them away for US$10.000. A Tesla Model 3 costs US$35.000. Elon stated they're aiming for a 25% gross margin, so each Model 3 would have a cost price of US$26.250.

So the DRIVE PX 2 would make up 38% of the cost - how is this possible? Could the remaining parts of the vehicle really only cost US$16.250?

 

[Tam]

...NVIDIA Drive PX 2 processing unit....

The article says it's

"40x more powerful onboard computer – Nvidia GPU Titan supercomputer"

So is Titan the same as Drive PX 2?

 

[calisnow]

Correct me if I'm wrong, but from what I've been picking up, the new autonomous capabilities in every new Tesla vehicle are powered by an NVIDIA Drive PX 2 processing unit. Each unit is liquid cooled supercomputer and as powerful as "150 Macbook Pros" (quote from NVIDIA CEO at CES 2016). BUT the list price of this unit is US$15.000!

The article says it's "40x more powerful onboard computer – Nvidia GPU Titan supercomputer" So is Titan the same as Drive PX 2?

You are wrong Frogwheels, but Nvidia and Tesla have not made it easy to figure this out. A few things to know:

1 - Drive PX2 is a development mule sold to automakers - it is not a product which will go into production vehicles (you have to click around a lot of articles to figure this out - most don't make this distinction clear). The production version of Drive PX2 is a small board which will have all the computation power of Drive PX2, but will cost only $100-$300 per unit if I recall correctly - but this chipset will not even sample for at least a year from now.

2 - Whatever Tesla is putting into cars today is based on Nvidia's GPU technology just like Drive PX2 is, and is much more powerful than what is in Teslas right now - but is NOT Drive PX2.

For all we know Tesla and Nvidia have built something custom for Tesla that is different from the Drive PX2 product Nvidia is selling to other automakers.

Drive PX2 seems like a bit more of an off the shelf product in any case. It is capable of learning to drive with zero human input and zero annotated images (ie Mobileye) - simply by following a video of someone driving a car and comparing steering wheel movement to what the camera sees out the windshield (this is the unsupervised learning that Mobileye's CEO likes to poo-poo at industry conventions). It can do sensor fusion as well.

Tesla, on the other hand, has been doing its own sensor fusion for two years now, has hired top chip designers and has a rock star team of AI and computer vision researchers. It seems reasonable to assume that whatever Nvidia built for/with Tesla is highly customized to Tesla's needs. Other than these items there isn't much else I know that is helpful, sorry.

 

[Frogwheels]

I see! That's excellent information, thanks for clearing this up. I did suspect that, but I couldn't find any articles stating that the PX 2 was only for development. Even in the NVIDIA CES presentation they made it sound like a Drive PX 2 was required in every car for autonomous driving.

That definitely clears up my question regarding the margin, thanks.

 

[calisnow]

BTW if someone has a more detailed technical explanation I am dying to hear it.

 

[Troy]

During the press conference, Elon said the system could run on different hardware and they have compared 3 different hardware systems, Intel, Nvidia and AMD and it was a close call between Nvidia and AMD but Nvidia's hardware was better. Also, the Nvidia Titan GPU is $1200 USD according to this article. I'm sure Tesla bought it for less than $800. The only thing Tesla confirmed was the Nvidia Titan graphics card.

 

[diamond.g]

During the press conference, Elon said the system could run on different hardware and they have compared 3 different hardware systems, Intel, Nvidia and AMD and it was a close call between Nvidia and AMD but Nvidia's hardware was better. Also, the Nvidia Titan GPU is $1200 USD according to this article. I'm sure Tesla bought it for less than $800. The only thing Tesla confirmed was the Nvidia Titan graphics card.

I wonder how much they down clocked it. The TDP for the Titan Desktop card is 250W which is a lot of heat to dissipate when running fill tilt, especially in the harsh environment of a car.

 

[MorrisonHiker]

I wonder how much they down clocked it. The TDP for the Titan Desktop card is 250W which is a lot of heat to dissipate when running fill tilt, especially in the harsh environment of a car.

Maybe that can route that heat to the passenger cabin on cold weather days!

 

[SageBrush]

I admit to being a little skeptical of this notion of Teslas being hardware ready for L4/5 as of today. Not only will sensors improve over time, but GPUs will too, and the SW will change to best use the new hardware. Granted, having a sensor on a side where one did not exist before is a fundamental improvement that is not going away, but the implied pitch that hardware today is future proofed strikes me as incorrect.

Dont get me wrong, I am elated at the Tesla approach to AP 2.0 and it has put me back on course to buying an M3 for sure. I just think Elon is setting unreasonably high expectations for these Gen1 AP 2.0 cars.

 

[maximus16]

We also don't know which version of the Titan it is, the current generation or previous.
Edit: I doubt this, but Nvidia and Tesla could of partnered up on a custom version of the Titan. Never heard of the Titan referred to as a 'Super Computer" like Musk has called it.

 

[diamond.g]

We also don't know which version of the Titan it is, the current generation or previous.
Edit: I doubt this, but Nvidia and Tesla could of partnered up on a custom version of the Titan. Never heard of the Titan referred to as a 'Super Computer" like Musk has called it.

Drive PX 2 uses Pascal, whereas Drive PX uses Maxwell. So it should be the current generation of Titan.

 

[LargeHamCollider]

The article says it's

"40x more powerful onboard computer – Nvidia GPU Titan supercomputer"

Which article says this? Not the original, Musk did say 12 trillion calculations per second by which I assume he means 12 trillion floating point operation per second or 12 TFLOPS. nVidia's latest Titan X is rated at 11 TFLOPS but could definitely be overclocked to hit 12.

I'm really interested to see if this actually is nVidia's monster 471mm^2 12 billion transistor chip.

 

[Tam]

Which article says this?

The link is at the red Titan .

 

[maximus16]

Drive PX 2 uses Pascal, whereas Drive PX uses Maxwell. So it should be the current generation of Titan.

If I am not mistaken.... These are different products. Drive PX2 is not what is going in the new cars, its just a standard GPU.

 

[LargeHamCollider]

The link is at the red Titan .

Well, that in combination with Musk's "12 trillion calculations/s" statement more or less confirms that it is nVidias massive GP102 die, 471mm^2 and 12 billion transistors is enormous in the chip world.
The thermal design power of GP102 is 250w, it's not going to pull that constantly but will still probably increase power consumption by 1-2w/mile at highway speeds, not too bad. I wonder if they have dual GP102s for redundancy...

 

[diamond.g]

If I am not mistaken.... These are different products. Drive PX2 is not what is going in the new cars, its just a standard GPU.

I actually think they are using the Tesla brand cards and not the Geforce brand. But that could just be semantics. Realistically they are either using the GP102 or the GM200.

 

[ecarfan]

Which article says this? Not the original, Musk did say 12 trillion calculations per second...

Yes he did say that, and he also said in the press call that the new vehicle computer is "40 times" more powerful than the previous one. Listen to the call.

 

[scaesare]

Correct me if I'm wrong, but from what I've been picking up, the new autonomous capabilities in every new Tesla vehicle are powered by an NVIDIA Drive PX 2 processing unit.

Each unit is liquid cooled supercomputer and as powerful as "150 Macbook Pros" (quote from NVIDIA CEO at CES 2016).

BUT the list price of this unit is US$15.000!

NVIDIA CEO Jen-Hsun Huang seems to be an admirer of Elon, so let's say he grants Tesla a discount and gives them away for US$10.000. A Tesla Model 3 costs US$35.000. Elon stated they're aiming for a 25% gross margin, so each Model 3 would have a cost price of US$26.250.

So the DRIVE PX 2 would make up 38% of the cost - how is this possible? Could the remaining parts of the vehicle really only cost US$16.250?

Actually, what I think you are referring to is the DGX-1. This is the large expensive rack unit that the neural-network learning/development/definition is done on. The resultant neural network defined set is then run in real time on much smaller-platforms, such as the DRIVE PX/PX 2 platforms, or whatever NVIDIA hardware Tesla is implementing.

 

[J1mbo]

40x more powerful than the Tegra 3 chip, or than the EyeQ3 chip?

 

[scaesare]

You are wrong Frogwheels, but Nvidia and Tesla have not made it easy to figure this out. A few things to know:

1 - Drive PX2 is a development mule sold to automakers - it is not a product which will go into production vehicles (you have to click around a lot of articles to figure this out - most don't make this distinction clear). The production version of Drive PX2 is a small board which will have all the computation power of Drive PX2, but will cost only $100-$300 per unit if I recall correctly - but this chipset will not even sample for at least a year from now.

2 - Whatever Tesla is putting into cars today is based on Nvidia's GPU technology just like Drive PX2 is, and is much more powerful than what is in Teslas right now - but is NOT Drive PX2.

Do you have sources for the above that states the PX 2 stuff isn't what's actually implemented in cars