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

[Matias]

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.

DGX-1 costs 129k USD.

 

[scaesare]

Mea culpa. I believe I was mistaken regarding the DGX-1.

I mistakenly read your post Frogwheels, and jumped to the wrong conclusion. Apologies for the bad info.

 

[AudubonB]

Uh oh - a score of posts chiming in on the OP's question, encompassing approximately a half dozen answers moderately to absolutely contradicting each other.

PLEASE....if those of you who think you know what the most probable answer/s is/are...summarize as best you can.

***A self-learning, self-correcting process. Just like something else in the real world, donchaknow?

 

[Frogwheels]

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.

Nope, I actually did mean the DRIVE PX 2. I didn't know about the DGX-1 though, so now my confusion is complete

Actually, after reading the article on anandtech.com where the Titan and Drive PX 2 are mentioned, I now again believe both the Titan X and the DRIVE PX 2 are intended to be used in consumer cars, with the DRIVE PX 2 being the more powerful option (6x the power of Titan X).

After all, if only the neural network training requires supercomputers, why not just feed them the recordings of sensor data back in the office? How come did NVIDIA then make an effort to design the DRIVE PX 2 to be able to be powered and fit into cars, especially electric cars?

E.g. Volvo will use 100 DRIVE PX 2 units:
NVIDIA's Deep Learning Car Computer Selected by Volvo on Journey Toward a Crash-Free Future

But wccftech.com (amongst other sites) mentions the DRIVE PX 2 has a price tag of US$15.000, which is just crazy.

So it seems Tesla picked the cheaper, less powerful Titan X, probably feeling that the Tesla Neural Net doesn't require a more powerful, expensive processing unit for proper functioning.

Here Jen-Hsun mentions the Titan X GPU being used to run a NVIDIA DriveNet neural network at 50fps:

At 6 mins in he also mentions that the Deep Neural Net will ultimately run on the DRIVE PX 2.

 

[McRat]

Correct me if I'm wrong, but the nVidia heavy lifter chip right now is Pascal? 10.6 trillion floating point operations per second in 32 bit mode? Dual chips is 20 TFLOPS.

Yeah, that would be the world's fastest supercomputer until the year 2002. The record in 2000 was 7 teraflops.

The silicon price won't be the issue. The software and engineering integration will. I doubt the silicon will be $1000 including the RAM.

 

[LargeHamCollider]

Frankly labels like PX2.0 DGX whatever are more or less irrelevant, what matters is the silicon, the silicon is almost certainly a GP102 die, the same that is used in the latest Titan X.

 

[Matias]

"As we reported, the system was expected to run on Nvidia hardware and while the company recently launched a few platforms built especially for self-driving cars, Tesla went with a less expensive solution: a Nvidia Titan GPU."

A look at Tesla’s new Autopilot hardware suite: 8 cameras, 1 radar, ultrasonics & new supercomputer

 

[Matias]

Is it this?

NVIDIA TITAN X Graphics Card with Pascal Architecture

 

[calisnow]

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

Yup and you are welcome to go research them yourself - there are articles stating what I just said - that the PX2 production hardware doesn't sample for another year. I for one am too lazy to go find them again.

 

[Matias]

Interesting info in electrek article "Since we are on retrofits, Musk did say that the new vehicles will eventually be able to upgrade the new onboard Autopilot computer since the access has been made relatively easy."

Is Tesla considering the possibility, that the new hardware is not enough for level 5?

 

[Tam]

Is Tesla considering the possibility, that the new hardware is not enough for level 5?

Historically, Tesla has been on the move all the time so I wouldn't be surprised if this is not a last hardware revisions.

 

[ORB]

From what I gathered, AP1 and now AP2 runs on a separate computing platform from that used for the driver UI (navigation, media etc.). So we should not expect any improvement in the UI responsiveness (especially navigation map redraws) from the original 2012 cars such as mine until that original Tegra chipset for the UI is upgraded someday? I hope that perhaps a future UI software upgrade that uses CPU resources more efficiently would help.

Corrections welcome.

 

[flashflood]

And if that's not confusing enough, NVIDIA has a product line called Tesla, including a model P100 -- completely unrelated to Tesla cars.

 

[LargeHamCollider]

Is it this?

NVIDIA TITAN X Graphics Card with Pascal Architecture

Yep

 

[Tam]

I thought it's a Titan as announced by Elon Musk. Now, nVidia says Drive PX2 Technology. What's going on?

Tesla Self-Driving Car Built on NVIDIA DRIVE PX 2 | NVIDIA Blog


Tesla Motors’ Self-Driving Car “Supercomputer” Powered by NVIDIA DRIVE PX 2 Technology


Tesla Motors has announced that all Tesla vehicles — Model S, Model X, and the upcoming Model 3 — will now be equipped with an on-board “supercomputer” that can provide full self-driving capability.

The computer delivers more than 40 times the processing power of the previous system. It runs a Tesla-developed neural net for vision, sonar, and radar processing.

This in-vehicle supercomputer is powered by the NVIDIA DRIVE PX 2 AI computing platform.

NVIDIA DRIVE PX 2 is an end-to-end AI computing system that uses groundbreaking approaches in deep learning to perceive and understand the car’s surroundings.

Our deep learning platform is open and lets carmakers first train their own deep neural networks on GPU supercomputers. Once loaded into the car, it processes the networks at high speed to provide the real-time, accurate response required for autonomous driving.

DRIVE PX 2 is in full production.

 

[Matias]

Apparently Drive PX 2 comes in multiple configurations

Autonomous Car Development Platform from NVIDIA DRIVE PX2

 

[J1mbo]

PX2 makes a lot more sense, as the chip is packaged for Automotive OEM use, unlike the TitanX card.

In the call, Elon mentioned AMD was a possibility. Not aware of any AMD DNN solutions that are packaged for automotive?

 

[diamond.g]

Yeah PX2 uses Pascal GPU and Denver CPU. Though the GPU config isn't entirely clear.

 

[LargeHamCollider]

Yeah PX2 uses Pascal GPU and Denver CPU. Though the GPU config isn't entirely clear.

Exactly, what this means is that PX2.0 uses the GP102 chip.

 

[electracity]

Interesting info in electrek article "Since we are on retrofits, Musk did say that the new vehicles will eventually be able to upgrade the new onboard Autopilot computer since the access has been made relatively easy."

Is Tesla considering the possibility, that the new hardware is not enough for level 5?

Yes! Leve;6! The car decides where you should be going, and takes you to that location. Parents everywhere rejoice.