Nvidia’s Drive PX Pegasus is a next-generation robotaxi brain

We string many words together around here describing the gaming performance of Nvidia's desktop and mobile graphics silicon, as well as the ever-expanding applications of the company's chips in the context of data center machine learning and artificial intelligence. The company wants to take that AI and put it out onto city streets with the Drive PX Pegasus, the next generation of its Drive PX platform for autonomous vehicles. Nvidia says Pegasus packs ten times the performance of its predecessor, the Drive PX2. The company wants its partners to deploy Pegasus in a future fleet of fully-autonomous taxis.

This is The Tech Report, so we'll pause a moment to talk about the hardware specifications. Pegasus delivers up to 320 trillion operations per second from two of Nvidia's Xavier SoCs coupled with a pair of discrete GPUs. The Xavier chips each feature embedded graphics hardware based on the company's Volta architecture. Chips based on the Volta architecture are only now spreading into well-funded datacenters, and the GPUs on the Pegasus board are based on Volta's future successor. Nvidia says those discrete chips contain specialized hardware for accelerating deep learning and computer vision algorithms. The company says all of this computing power will fit into a system the size of a license plate. Since the company made the announcement at GTC Europe in Germany, we aren't sure if that means a U.S. or a European plate, though.

The Pegasus is designed to receive an Automotive Safety Integrity Level D certification, the highest standard of functional safety for road vehicles as defined by the ISO. The system has a CANbus interface for connecting with existing automotive electronics and the next-generation Flexray automotive communications bus. Pegasus also has 16 dedicated inputs for cameras, radar, lidar, and ultrasonics, in addition to multiple 10 Gigabit Ethernet connections.

The green graphics guys say their system could allow people to reclaim billions of hours lost behind the wheel of a car. The company goes as far as saying vehicles built around Pegasus' technology could eschew steering wheels, pedals, or mirrors, thereby offering increased flexibility to automotive designers. Nvidia claims that machine-driven vehicles could eventually be safer than human-driven cars because AIs aren't distracted, tired, or emotional. The deletion of mechanical features like the steering column from the vehicle interior could potentially improve crash survivability, too.

Nvidia says the Drive PX Pegasus will be available to its automotive partners in the second half of 2018. If you don't work in the purchasing department of a major automotive manufacturer or supplier, the price is probably irrelevant to you. Nvidia says it has 225 partners working with its Drive PX computing platform and that 25 of those are developing fully-autonomous robotaxis.

Comments closed
    • psuedonymous
    • 2 years ago

    Zoom and enhance! Those ‘post-Volta’ GPUs sure look like they use discrete GDDR rather than on-package HBM. Could mean the Tensor Cores (and other Pascal/Volta changes like the FP64 cores splitting for FP32 and FP16 ops) will start working their way down from the Gx100 dies to the Gx10x dies.

    • phileasfogg
    • 2 years ago

    >>>>>> … and the next-generation Flexray automotive communications bus

    there is no “next-generation” FlexRay bus. F-Ray has been around for many years and will likely be eventually phased out in favour of automotive Ethernet (802.3bw = 100Mbps on a single-twisted-pair and 802.3bp = 1000Mpbs on a single twisted pair)

    • blastdoor
    • 2 years ago

    I’m a little skeptical that this will enable true level 5 autonomy, but I guess we’ll find out.

      • cygnus1
      • 2 years ago

      I kind of doubt it’s even necessary. Very much brute forcing the issue. [url=https://arstechnica.com/cars/2017/10/blue-roads-and-glowing-signs-how-this-startups-tech-lets-cars-see-the-world/<] This little startup has an interesting take on processing the inputs that won't need nearly as much compute hardware. [/url<]

        • blastdoor
        • 2 years ago

        Meh — I’m skeptical of that, too. Certainly it’s better to have hyper accurate maps than to not have them, but I think it would be a big mistake to build a system that *requires* such maps. Furthermore, I think it’s a mistake to build a system that requires a continuous Internet connection, or even a reliable GPS signal.

        My intuition is that the AI in the car needs to be pretty darned smart. It needs to be able to drive safely even if it’s cut off from the Internet and even if it has to rely on a combination of street signs and “normal” (i.e. slightly outdated) maps to get from here to there. And in a pinch, it should be able to get by without GPS (read street signs and addresses).

        Doing all that is something that a human can do. I think a level 5 car needs to be able to do the same thing. The reason I’m skeptical about Nvidia’s solution is that I doubt it’s powerful enough or customized enough. I suspect what’s needed is hand-in-glove co-development of software and highly customized silicon. But that’s just a hunch… I don’t know for sure.

          • cygnus1
          • 2 years ago

          I can definitely agree with all that. The sheer storage and bandwidth requirements of the methods that would need this kind of compute hardware make level 5 way over the horizon. Given how much smaller the data is from this startups mapping method, you could actually cache a relatively large area to the car and allow it to search the entire data set in a reasonable time to approximate it’s location based on what it can see, no need for any GPS or even cellular or wifi location tracking.

          The startup I linked to still has hyper-accurate maps though. And it can be updated on the fly and submitted for distribution to other cars, it just needs way less compute to do it. If their method proves to be just as good as the brute force method, the startups method will win just on an economic basis.

            • blastdoor
            • 2 years ago

            Hmm… I think I misunderstood what the startup is doing. Thanks for the extra info!

            Another thought that occurs to me is that Level 4 offers quite a bit in the form of a geofenced taxi service. It’s probably ok that Level 5 is way off in the future.

    • Neutronbeam
    • 2 years ago

    Did they mention anything about encryption to prevent hacking? Because that’s what concerns me most.

      • chuckula
      • 2 years ago

      Is hacking a big concern: Totally.

      Is encryption the solution to hacking: Not particularly.

    • ludi
    • 2 years ago

    Now combine this with the classic Dustbuster cooler and you have yourself both navigation and propulsion.

      • UberGerbil
      • 2 years ago

      You know, autonomous street sweepers are kind of an interesting idea. Though the noise, the noise….

        • Mr Bill
        • 2 years ago

        All the cats in the neighborhood riding them…

    • chuckula
    • 2 years ago

    Level 5 self-driving cars are great & all but if you are going to call it Pegasus then it better be able to [url=https://www.youtube.com/watch?v=wNXGcvHPqnI<]phase through solid matter to avoid collisions[/url<] Nvidia.

      • AnotherReader
      • 2 years ago

      I regret that I have just one upvote for that comment.

      • Dposcorp
      • 2 years ago

      I gave you a “upvote” but would have preferred if you linked Pretty Pretty Pegasus.
      [url<]https://www.youtube.com/watch?v=ufjlJQCWmrk[/url<]

      • Mr Bill
      • 2 years ago

      Begs the question, how rapidly can one move through solid matter with that cloak?

    • Helmore
    • 2 years ago

    Given the general shape of the PCB, if that picture is a lifelike representation, then I’d assume they’re talking about a US license plate. EU license plates are much longer in shape, not a square as this. Besides, a EU license plate it only about 28% bigger than a US one in surface area.

    • Vigil80
    • 2 years ago

    Societal acceptance of this technology is probably a long way off. But I’m ready for robocar to ferry me around while I watch Netflix, eat, or nap.

    I wonder how many municipalities here in the US will require an “attentive user” even after the cars themselves are proven and street legal.

      • Chrispy_
      • 2 years ago

      I’ve been in enough human-operated taxis driven badly that I would happily accept a ride in a robotaxi.

      At least they won’t be high/drunk/crazy/incompetent.

      • UberGerbil
      • 2 years ago

      The real question is one of liability: when accidents inevitably occur (and they’re not clearly and completely the fault of something other than the autonomous vehicle), who is at fault? The manufacturer? The outsourced software? The sensor, CPU, and other component makers? The vehicle owner?

      Having an “attentive user” who is not involved in guiding the vehicle, and then expecting that user to be able to take over suddenly in an anomalous situation and do anything to make the situation better, is a fool’s errand. At best they’re unlikely to do any better than whatever autonomous systems would do; more likely [url=https://hbr.org/2017/09/the-tragic-crash-of-flight-af447-shows-the-unlikely-but-catastrophic-consequences-of-automation<]they're going to make the situation worse[/url<].

        • Chrispy_
        • 2 years ago

        I get the feeling that in an accident involving an autonomous vehicle, the blame and legal/financial responsibility will be exceptionally easy to assign.

        Let’s face it, these things are sensing, reacting, and logging data from dozens of sensors hundreds of times a second, and their programming ultimately errs on the side of caution. If something goes wrong, it’ll be ludicriously easy to provide evidence that the other party was at fault, and they most likely are.

          • UberGerbil
          • 2 years ago

          While that’s true, there are always corner cases that don’t show up until some extremely unlikely combination of factors come together to produce a failure mode that no one anticipated or tested for. And that’s going to be especially true in the early days.

          And reality is such that there can be situations where there is [i<]no[/i<] good choice for the software to make, so when it makes a choice and somebody dies, they're going to get sued over the assertion that it should have made a different choice even if it was equally bad.

            • Chrispy_
            • 2 years ago

            If the failure mode wasn’t predicted and it’s found out that a human behind the wheel would easily have avoided the accident, the manufacturer will be sued. The good news for them is that they’ll have insurance against this (hopefully), and apply the learned lesson to all future and current vehicles via software updates.

            Human drivers on the other hand, never learn. In the US alone, 40,000 people are killed by drunk/high/tired/phone-using/unconscious drivers and 2.5 million are injured [b<]every year[/b<]. Insurance companies exist for the 40,000 and 2.5 million annual victims, I'm willing to bet they'd much rather insure predictable, sober, drug-free, non-distracted AIs than (statistically-proven) terrible human beings.

        • frenchy2k1
        • 2 years ago

        Couple of points: there are multiple “levels” defined for autonomous cars. The level targeted here, level 5, is the highest and the current end goal, where the car is fully autonomous. There will be no “attentive user”. Even if there is, there will be no control for that user to take. Those type of fall over to a possible driver are up to level 4 (level 2 is basically lane assist, level 3 requires an attentive user with fall-over within a few seconds, level 4 is fall over within a minute, as the car can predict when its programming is insufficient).

        Discussion about liability has started and so far, manufacturers of autonomous vehicles have accepted to shoulder the responsibility (certainly for level 5). If a component failed in a manner out of spec, they may then turn to their suppliers (this is a normal liability chain).

        It will take time to reach all the legal framework and technical competency, but we’ll get there. Remember, we have over 30k+ death on the road in the US alone every year. Competent self driving cars can only help and once that stage is reached, it will only be a matter of time before manual driving is outlawed for safety reasons.

    • iatacs19
    • 2 years ago

    “Since the company made the announcement at GTC Europe in Germany, we aren’t sure if that means a U.S. or a European plate, though.”

    LOL, well played! 😀

      • Wirko
      • 2 years ago

      The “European plate” is not the same size across all countries.

      Heck, it’s not necessarily the same size across a single car!
      [url<]https://it.wikipedia.org/wiki/Targhe_d%27immatricolazione_italiane#1951.E2.80.931976[/url<] (these big+little plates are still valid in Italy)

      • chuckula
      • 2 years ago

      Look, there’s no way that a 5 ounce plate is going to carry a 320 Tflop autonomous driving controller.

        • UberGerbil
        • 2 years ago

        Well, as long as there isn’t an air gap between whatever TIM is on the plate and the controller, you won’t have to worry about overheating when you overclock the controller to make it drive even in better.

        • Mr Bill
        • 2 years ago

        More troubling than encryption to prevent hacking; would be somebody stealing the license plate. Although, it would probably run pretty cool.

        • Mr Bill
        • 2 years ago

        That depends. What is the airspeed of an unlaiden 5 ounce plate?

          • Chrispy_
          • 2 years ago

          African or European?

        • Chrispy_
        • 2 years ago

        Supposing two plates carried it together, on a line held under the dorsal guidance system?

Pin It on Pinterest

Share This