Tesla using NVIDIA DRIVE PX 2 as autonomous driving controller
21 October 2016
Tesla Motors announced on Wednesday that all Tesla vehicles—Models S, X, and 3—will be equipped with an on-board “supercomputer” that can provide full self-driving capability. (Earlier post.) This in-vehicle supercomputer is powered by the NVIDIA DRIVE PX 2 AI computing platform. (Earlier post.)
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.
NVIDIA DRIVE PX 2 is an end-to-end AI computing system that uses advanced approaches in deep learning to perceive and understand the car’s surroundings.
The NVIDIA 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.
This is a big scoop for NVIDIA. Will be certain to attract new customers other than Tesla. That Tesla has chosen to use it means it is the best on the market currently for that type of computing jobs. NVIDIA's CEO also drives a Tesla. Bet he ordered a new one with this tech.
Posted by: Account Deleted | 21 October 2016 at 03:27 AM
Electrec has a story that the NVIDIA PX2 sell for 15.000 USD a piece to OEMs in low volume like 10 per year for Audi and 100 per years for Volvo. Tesla buying at least 8000 per month should get it at a much lower price but it will still be costly for sure. It also use 250 watt witch is a lot. However, this level of computing power is needed for full autonomy AI autopilots so there is no alternative. And it will cheaper for sure.
https://electrek.co/2016/10/21/all-new-teslas-are-equipped-with-nvidias-new-drive-px-2-ai-platform-for-self-driving/
Posted by: Account Deleted | 21 October 2016 at 11:34 PM
Let's hope that the NVIDIA "supercomputer" and software have improved significantly over the graphics cards foisted on Apple owners that overheated and crashed systems.
Crashing a computer is inconvenient, but crashing an autonomous driving system can be deadly.
Posted by: DL | 24 October 2016 at 12:08 AM
It is liquid cooled but the cooling system needs to be 100% reliable. Maybe they have two pumps for that just in case. It consumes 250 watt so it will quickly break if the cooling pump stops.
The devise is made solely for automotive use so I assume it is designed to be a lot more reliable than a PC for consumer use in all regards.
Posted by: Account Deleted | 24 October 2016 at 12:24 AM
DL Tesla only buy the PX2 from NVIDIA. The software is Tesla made. Their may be driver and firmware that NVIDIA supplies but the AI pilot software is 100% Tesla made. I think Tesla has a deliberate strategy of not making it easy for others to copy Tesla so if Tesla can make it themselves in volume they will. Like the AI pilot software and the new camera sensors that are Tesla made also. Tesla's data center is also Tesla made. They will start using a new battery cell that is Tesla unique. It will only be made in that form factor with that chemistry on the Giga fac. The falcon doors are Tesla made etc. It is actually really hard to make a Tesla car if you are not Tesla. Tesla will enable a lot og growth by being the only maker of a fully autonomous car for a few years ahead of everybody else so they do it alone.
Posted by: Account Deleted | 24 October 2016 at 01:06 AM
For safety reasons, sensors, controllers, cooling and computers on ADVs will have to be double just like aircraft autopilots are?
Posted by: HarveyD | 24 October 2016 at 10:45 AM
If Tesla is so far ahead of the rest of the players as a certain other person thinks they are then they'll be the ones to find out the hard way.
FMEA. MTTFd. They mean something. I'm not convinced Elon Musk knows or pays heed. Everyone else who designs stuff for the auto industry has nightmares about those. Ignoring them doesn't make them go away. It just ensures that you'll get bitten by them.
And just for the record, I'm not anti-Tesla. A customer of mine builds parts for them. I love a lot of the things they've done ... but not Autopilot, and not falcon-wing doors, both of which I consider to be FMEA failures!
Posted by: Brian P | 24 October 2016 at 07:30 PM
Brian Tesla’s cars best every other car on the market in terms of safety. They got top ratings at all the authorities that are testing for that. Their Autopilot is approved by all authorities jurisdiction by jurisdiction before it is deployed feature by feature just like the autopilots of other car makers. However, all reviews you can find about autopilot systems say Tesla’s implementation is the best by far. Also over 95% of those that order a Tesla order it with the autopilot option.
About Falcon doors. They appear to work as intended and not break. There has not been any recalls yet. However, more years need to pass before we know whether they are as durable as regular doors. In most ways the Falcon doors are progress because it is easier to get in and out and they do not require much space or get in the way when opened. But they also cost more, open less fast and open the car more so rain and snow can enter easier when opening in bad weather. Tesla IMO should therefore also make a version with regular doors. I think Tesla will reach more customers that way by offering both solutions to the doors.
Posted by: Account Deleted | 25 October 2016 at 12:15 AM
“Tesla’s cars best every other car on the market in terms of safety. They got top ratings at all the authorities that are testing for that.”
Nonsense.
Post one (just one) affirmation by ANY Government or Industry rating consortium that states in ANY form that Tesla cars “best every other car on the market”. No such assertion exists. Taking just Euro NCAP Crash Testing results for “Large and Executive Sedans” 2012-2016 as examples, the 2014 Tesla Model S finished behind the 2012 BMW 3 series, among about a dozen others, in both Adult and Child rating.
The most telling metric of safety (or lack of it) is fatalities in US highway data, where Tesla is shaping up very poorly. Using the fatalities per million mile metric (Model S: 330) that Musk loves to quote, the Model S lags the following examples:
Chevy Impala: 343
Honda Accord: 632
M/B C Class: 1200
Lexus ES 350: 1333
BMW 328: 1714
Tesla’s “superior” safety is not true.
Posted by: Herman | 25 October 2016 at 02:43 PM
Correction to above: not "fatalities per million miles" but rather the inverse: million miles per fatality.
These are from Insurance Institute for Highway Safety data for all but Tesla; Tesla numbers are based on actual fatalities data records and a projection of miles driven. 330 Mmi/f is higher than Tesla's own assertion, BTW.
Posted by: Herman | 25 October 2016 at 03:13 PM
Autopilot is not "approved by all authorities jurisdiction by jurisdiction before it is deployed feature by feature just like the autopilots of other car makers". That ain't the way it works (for now). In North America, there is no third-party approval body for ANY such thing. The auto manufacturers self-declare compliance with FMVSS (USA) or CMVSS (Canada); NHTSA or Transport Canada review submitted paperwork, do occasional spot-checks, and investigate complaints. That's it. Some other jurisdictions (notably EU) do have a third-party type-approval process, but the approval bodies have a set of rules that they follow and check compliance with, and if it conforms to the standard then they grant WVTA (whole vehicle type approval). There currently IS NO standard that governing bodies apply to driver-assistance systems, never-mind self-driving systems! If there is no standard then the type-approval body has no grounds to say "Nein"!
Those falcon-wing doors have been troublesome, and it doesn't take a whole lot of searching to dig up all manner of issues with them. Check Consumer Reports, if nothing else. Oh look; they are skeptical of Autopilot, and the Model X is not recommended due to the falcon-wing doors: http://www.consumerreports.org/car-reliability/car-reliability-survey-2016/
Posted by: Brian P | 25 October 2016 at 05:59 PM
I should also add that with regards to the "million miles per fatality" statistics, when you get towards the vehicles that have big numbers, it's because they have a small number of fatalities. This, of course is a good thing, but it also means that the sample size (of fatalities) is small, and random situations that lead to just one more or one fewer fatality can shuffle the order with no real underlying reason behind it. The Accord and Impala numbers should have good statistical significance, but the others could be in any order depending on random circumstances.
"Million miles per fatality" is also strongly affected by demographics. Two vehicles with identical collision performance, but one of them is bought by teenagers and another by fortysomethings, will have drastically different real world crash frequency. There were a number of minivans a couple vehicle generations ago which had notoriously poor offset-frontal crash performance, but yet few real-world fatalities, because of who drives them and how.
I'm not concerned with Tesla's collision safety.
Posted by: Brian P | 25 October 2016 at 06:12 PM
No dispute with your overall points, Brian. But two important points here:
(1) the demographics of the high-end Lexus SUVs, M/B large sedans, and BMW 5-ser and up are likely very similar to Tesla S and yet their fatality statistics are far better across the board. I recognize that their stats have more statistical stability owing to many more registered vehicle miles, but nonetheless this is not without note.
(2) I'm not "concerned" with Tesla safety, either; I would not be uncomfortable driving or riding in one. Look again at Henrik's remark: "Tesla’s cars best every other car on the market in terms of safety. They got top ratings at all the authorities that are testing for that." My point was to establish both a lack of regulatory or industry consortia endorsement of such a thing, and to exhibit statistics as they exist today that reflect Tesla is in fact far from "besting every other car."
Posted by: Herman | 25 October 2016 at 09:04 PM
When things are taking out of context the truth is distorted which is not fair to Tesla. The reliability issues with Model S and X that consumer report has noted is only for the initial production cars. As for any automaker and especially those that are making entirely new and unique products like Tesla it will take some time (typically a year) before the production is nearly error free. It is not a Tesla customer problem at all as these issues are fixed under warranty and Tesla customers are known to be more satisfied with their cars than any other car buyers on the planet. The acid test is customer satisfaction and Tesla has better ratings on this one than all of its competition.
The other out of context thing is that I was referring to safety by the safety test performed by the authorities. Tesla does indeed best all its competition on this one also see link below fx.
Herman you start to through completely undocumented data up that refer to death per million miles not safety ratings. Unless you provide sources for these data and circumstances they were collected etc it is worthless.
It is true that Musk has referred to the death per million mile data. So far, only one death on Tesla autopilot for now well over 200 million miles possibly 300 million miles. The specific accident will most likely not happen again as the new Autopilot version 8 software is using the radar for main navigation and not the camera that was blinded in the death accident. The accident itself was a freak accident with a guy that in all likelihood was watching a video while he was supposed to drive. Again with version 8 he could probably not have been that much inattentive to driving as the version 8 software constantly prompts you to keep the hands on the wheel and deactivate autopilot if this is ignored.
The new hardware version 2 autopilot in all Teslas produced going forward will be much more safe in the situations that Hardware version 1 could self-drive but it will also drive itself in a lot of new and much more challenging circumstances so the overall death per million mile will probably stay the same or about 1 death per 200 million miles which is over 3 times as good as the global stat and over 2 times as good as the US stat.
I take note that none of you would be worried about driving a Tesla.
http://www.businessinsider.com/tesla-tesla-model-s-achieves-best-safety-rating-of-any-car-ever-tested-2013-8?r=US&IR=T&IR=T
Posted by: Account Deleted | 26 October 2016 at 01:35 AM
"Tesla does indeed best all its competition on this one also see link below fx."
Nonsense again. Good grief.
The NHTSA clearly stated, after Tesla published this assertion, that the use of crash data in the fashion claimed by Tesla is misleading. There is no such thing as "5.4 stars", and in fact the interpretation of the data as Tesla did assumes that the data can be statistically parsed to indicate a "better" star rating. It can't. Tesla was specifically told to halt these assertions and in fact NHTSA set in place a policy to tighten the public use of their data by manufacturers because of Tesla's incorrect assertions.
Tesla got five stars across the board. That's great. It's not an easy rating to get. But it's not unique. The 2016 Nissan Maxima. The 2017 Acura TLX. The 2017 Honda Accord. Two versions of the 2017 Honda Civic... and others.
BTW, the crash standards to achieve 5 stars were made tougher at the end of 2015 with the addition of the frontal oblique test. Tesla hasn't taken this test yet.
Tesla is NOT recognized by ANY regulatory or industry consortia as the safest car. No such legitimate rating exists, regardless of how Tesla likes to invent such a rating (clearly denied by the agency) and how much Tesla advovcates want to desperately believe it.
In fact European testing and IIHS fatality data would strongly suggest otherwise.
If you want to argue the data, use the data. If you want to argue regulatory agency opinions, use the official statements of regulatory agencies.
Tesla cars are not "unsafe". But to call them "safest" is nonsense.
Posted by: Herman | 26 October 2016 at 06:16 AM
Still no sources Herman. And Tesla's safety score is still the highest ever recorded. However, Tesla is not allowed to say that in public for some reason probably old-auto lobbying. Tesla crumble zone is bigger than any other car on the market and it does not carry flammable liquids. Of cause it is safer. It also does not kill numerous people through air-pollution like all the gassers that kill millions every year prematurely on that account.
Posted by: Account Deleted | 26 October 2016 at 06:32 AM
"Herman you start to through completely undocumented data up that refer to death per million miles not safety ratings. Unless you provide sources for these data and circumstances they were collected etc it is worthless."
Source of data is IIHS.org, which I have used with a heavy dose of conservatism to Tesla's benefit, including
(1) a correction to the raw fatality data for the six US fatalities (reduced to five) to correct for demographics in a manner that's probably more lenient than IIHS.
(2) no adjustment for the very large number of fleet Impalas and the high mileage logged
Again: as Brian stated the big problem with Tesla data is that the car has only about 150,000 US registered vehicle years of exposure (which again is a conservative estimate IN FAVOR of Tesla that makes its fatality data better) which means that any death swings the data hard. The real numbers without adjustment aren't as good as 330 Mmi/f; they're worse.
But in the context of newly introduced cars with little exposure there is precedent. Here is some information on the following models from 2009-2012:
Audi A4 AWD, Honda Odyssey, Kia Sorrento 2WD, Lexus RX350 4WD, MB GL-Class AWD, Subaru Legacy AWD, Toyota Highlander 4WD, Toyota Sequoia 4WD, and Volvo XC90 4WD.
Adjusted for inflation, all but the top-of-the-line MBs are less expensive cars than the Model S. Except for the Lexus, none had drivers assist features. Between all of them they have clocked over 900,000 registered vehicle years as of Jan 29 2015. Assuming 12,000 miles per registered year, the total miles driven is 10.8 Billion.
Now: based on data gathered by IIHS members and thorough review of all the death statistics submitted by manufacturers, here are the death statistics (within the confidence levels that any large data set can provide) for all of these models, according to the IIHS ---
Multi-vehicle crash death rate: zero
Single-vehicle crash death rate: zero
Rollover death rate: zero
Tesla numbers thus far are not good -- not terrible, but not good. It's NOT unsafe. I would fall asleep in the quiet cabin as a passenger without concern (as long as the driver isn't doing a hands-off Autopilot stunt). But there are safer cars, based on many, many statistics available to you and anyone who wants to find them.
Posted by: Herman | 26 October 2016 at 06:46 AM
Herman thanks for the link to IIHS.org . I have bookmarked it. It does not have Tesla in it so currently it is not good for any comparisons. I am a data analyst so I know better than most that 1 fatality is no statistic for making conclusions at all. We need billions of miles and at least 30 deaths before it make sense to discuss safety in scientific terms. All we can do is wait for that to accumulate. Nevertheless Model S and Model X has better crashworthiness than any other car ever build. All test by authorities has proved that. On top of that Tesla is making a fully autonomous autopilot that ultimately are going to be ten times less likely to crass the car than a human driver in all circumstances. It is hell for old-style automakers as they make most of their profits on selling new cars and spare parts to fix traffic accidents. They do not want self-driving cars to happen but they now know it is inevitable and that they will bankrupt if they do not do it themselves. Cars without fully self-drive will have zero demand from 2025. Cars will still be sold in 2025 that have manual controls so you can drive them yourself if you want to. However, the law will require that the cars come with an autopilot that takes over if it detects an accident is about to happen as a result of human driving behavior. Full autopilot will be a mandatory safety feature for all new cars just like brakes and safety belts.
Posted by: Account Deleted | 27 October 2016 at 12:24 AM