Better have right driver for that speed 😀

sverek:

Better have right driver for that speed 😀

In this case it is left driver 😉

sverek:

Better have right driver for that speed 😀

Well it is a GeForce experience 🙂

Hilbert Hagedoorn:

Well it is a GeForce experience 🙂

Just wait 'til they turn RTX on 😉

m4dn355:

In this case it is left driver 😉

Wait till you see the hotfix driver.

lucidus:

Wait till you see the hotfix driver.

No reason to worry. If anything ever goes wrong... => Exciting Visual Effects => Adrenaline Experience to Die For

OK I'm missing something. Why do we need cars that drive themselves? And would you get on a plane flown, and hopefully landed, only by a computer?

NewTRUMP Order:

OK I'm missing something. Why do we need cars that drive themselves? And would you get on a plane flown, and hopefully landed, only by a computer?

I would not mind fully automated car where I just enter destination. Party all the way.

500W and they call this energy efficient? As for "This is not a demo, this is something you can get right now" ok, so where can I get this? Why is this dubbed a test? Why is still obviously in development? Look, Nvidia, this work you've done is cool and as far as I can tell, it's very well done. But don't up-sell it's capabilities and progress for more than it really is. Tell us facts, not your (Nvidia's) own personal opinions.

NewTRUMP Order:

And would you get on a plane flown, and hopefully landed, only by a computer?

Uh... planes are mostly flown by a computer, or at the very least, capable of being flown by one. And for the most part, I would trust a computer to fly and land a plane over a human.

schmidtbag:

500W and they call this energy efficient? As for "This is not a demo, this is something you can get right now" ok, so where can I get this? Why is this dubbed a test? Why is still obviously in development? Look, Nvidia, this work you've done is cool and as far as I can tell, it's very well done. But don't up-sell it's capabilities and progress for more than it really is. Tell us facts, not your (Nvidia's) own personal opinions.

500w power envelope is energy efficient for an automated drive system. Have you ever seen comparable ones, like tesla? Much more power efficient than those considering capability. Lastly, 500w is nothing. Alternators in cars have so much overhead and it's simple enough to install a higher amperage alternator if need be. 150amp alternator x 14 volts is over 2kw. That's practically free energy.

Meh! i live here (Silicon Valley) and the simple fact is Nvidia wanted a piece of the pie after their gpus were used by Google/Waymo/Uber cars. who could blame them as this is a huge emerging market, especially for electric self driving cars. and someone asked why? there are several whys; Safety, denser traffic flow and ultimately, the end of auto maintenance by having electric self drivers cheaper than driving yourself (incl. insurance, fuel, wear and tear, oil changes, tires, etc...) of course this is an urban model, but hey that's where the vast majority of people live in every country.

Agent-A01:

500w power envelope is energy efficient for an automated drive system. Have you ever seen comparable ones, like tesla? Much more power efficient than those considering capability. Lastly, 500w is nothing. Alternators in cars have so much overhead and it's simple enough to install a higher amperage alternator if need be. 150amp alternator x 14 volts is over 2kw. That's practically free energy.

A few things: 1. No, 500W really isn't efficient, especially if we're talking electric cars, where they already have a range issue. 2. Have you seen what Tesla uses? Because their system is based on the Drive PX2, a 125W system. That's perfectly acceptable, considering the autonomous driving capabilities of those cars. 3. Electric cars don't use alternators. But even if we're talking combustion engines, no, it's not free energy. In case you're not aware, the more power that is drawn from an electric generator (like an alternator) the more mechanical resistance it puts on the system. This resistance will decrease fuel economy.

Agent-A01:

500w power envelope is energy efficient for an automated drive system. Have you ever seen comparable ones, like tesla? Much more power efficient than those considering capability. Lastly, 500w is nothing. Alternators in cars have so much overhead and it's simple enough to install a higher amperage alternator if need be. 150amp alternator x 14 volts is over 2kw. That's practically free energy.

There is no such thing as free energy. Take DC electric motor (call it alternator if you want), rotate it by mechanical source of energy. There is voltage induced due to coils traveling through static magnetic field. If you short circuit electrical power outputs, there will be equivalent breaking force applied to mechanical rotation input as is Voltage * current. In other words: Alternator which has absolutely no load applies breaking force to mechanical source of energy equal to friction in it and rotating feromagnetic metal force resisting static magnetic field. Which is tiny force. Byt every real watt drawn from alternator is breaking force applied to its mechanical source.

just a point about reality... whether or not a politician wants to use the words "climate change" or "global warming" doesn't matter. what matters is the reality of actual water levels, storms, evaporation rates (esp. for the air forces) and the related temperature levels. this is why the U.S. military has been dealing with climate change, while the current politicians in charge are nay-sayers. the national security of every nation is at risk literally. and honestly, does anybody really think over 500 million extra cars that never existed before 2003 and hundreds of coal burning power stations that never existed before do not have an effect on a Closed System called planet earth? and that's just the Chinese/ Indian cherry on top of our sins since 1918.

oh yeah the point was that self driving electrified cars are the actual future

schmidtbag:

A few things: 1. No, 500W really isn't efficient, especially if we're talking electric cars, where they already have a range issue. 2. Have you seen what Tesla uses? Because their system is based on the Drive PX2, a 125W system. That's perfectly acceptable, considering the autonomous driving capabilities of those cars. 3. Electric cars don't use alternators. But even if we're talking combustion engines, no, it's not free energy. In case you're not aware, the more power that is drawn from an electric generator (like an alternator) the more mechanical resistance it puts on the system. This resistance will decrease fuel economy.

Wasn't the PX2 a 250W system? Unless there are multiple versions they use that I don't know about. Anyways, I said it is considering the capability. Tesla markets their cars as semi-autonomous; they require a driver to assist the vehicle due to multiple reasons. They are not fully-autonomous. To produce a 100% driverless vehicle would require a massive amount of more processing power to ensure there are no mistakes then one that is 'almost' autonomous. Said vehicle would need many more data inputs such as more cameras, LIDAR, ultrasonic sensors and so on; the current PX2 would not have the capability to process the necessary data to make decisions. Teslas still make mistakes, quite a few crash videos of those so it's clear the system is not powerful enough yet. Of course wattage is more important for fully-electric cars and every bit helps range, but the presumption was that the Ford in the photo is a gas-driven vehicle. There are plenty of gas/hybrid semi-autonomous vehicles. As for alternators, they are practically free, in context. An additional few hundred watts on a belt-driven alternator is extremely minor. A couple hundred watts of headlights on vs off will hardly affect MPG. You don't see people complaining that their MPG sucks at night or when they turn on their stereo system do you? It's because even combined, the difference is negligible. So considering everything, 500w is a lot in principle; but that system is efficient considering semi-autonomous is not directly comparable to a fully-autonomous vehicle. and lastly, if you compare TOPS performance, you will see that the 500w system is many times more energy efficient(perf/watt)

https://thumbs.gfycat.com/SnarlingAthleticAfricanbushviper-size_restricted.gif

Agent-A01:

Wasn't the PX2 a 250W system? Unless there are multiple versions they use that I don't know about. Anyways, I said it is considering the capability.

There are multiple versions; the Tesla version uses less power. And even then, 250W isn't that bad. Its still high but not "this may need liquid cooling" high. As for the capability, you're getting into a bit of a gray area here because that tends to be more up to the software rather than the hardware. As far as I'm concerned, Nvidia is just accomplishing their goal without any optimizations. I'm sure it's handling more sensor data than it needs, I'm sure the code could be written more efficiently, and as a result, they don't need that much processing power. I'm not faulting them for that specifically; as far as I'm aware, they haven't been doing this for as long as other organizations, so they're making good progress.

Tesla markets their cars as semi-autonomous; they require a driver to assist the vehicle due to multiple reasons. They are not fully-autonomous.

Yes, and many of those reasons are due to experimentation and liability issues. They're still adding new features to it. Take Nvidia's setup in a European city or US east-coast and I'm confident they too would go for the "semi-autonomous" label. Driving in Silicon Valley isn't a real challenge; there have been successful attempts at full autonomy using weaker hardware several years ago in southern CA.

To produce a 100% driverless vehicle would require a massive amount of more processing power to ensure there are no mistakes then one that is 'almost' autonomous. Said vehicle would need many more data inputs such as more cameras, LIDAR, ultrasonic sensors and so on; the current PX2 would not have the capability to process the necessary data to make decisions.

Although I would agree that there is quite a considerable amount of complexity between a "mostly autonomous system" vs a "fully autonomous system", I don't agree that there is a correlation to that when compared to sensors and processing power. An attentive human driver can drive at least as good as existing nearly-fully-autonomous systems, and yet all we have (when compared to digital hardware) is a stereoscopic ~200 degree camera, some mirrors, and a couple of microphones. Just having a 360 degree LIDAR with total precise control over the vehicle is already a HUGE advantage over human capabilities and is not that computationally taxing. The real challenge in self-driving AI is understanding how to respond to the ever-changing variables of the environment. Of course, some other sensors are necessary in order to add more useful information that something like a LIDAR can't observe. For example, transparent objects or highly reflective materials, which can give false readings. Those would warrant a need for other "failsafe" sensors, like ultrasonics. But, just simply adding more sensors in general isn't going to solve the problem, it just adds more data to be processed. Meanwhile, the more complex something is, the greater chance there is that it will fail. tl;dr: There's a steep level of complexity between currently available technology and a truly fully-autonomous system, but, that does not mean we need over 4x the processing power to do so. What we need is an improvement in logic, and that does not have a direct correlation with processing power. Just to clarify, I do agree more processing power is necessary, my only point is the level Nvidia went to seems inefficient. I don't doubt they did a good job with their AI, I'm just saying it's inefficient.

Teslas still make mistakes, quite a few crash videos of those so it's clear the system is not powerful enough yet.

And how many of those accidents are due to human drivers? How many of those accidents are due to failed sensors? How many of those accidents are due to programming errors? How many of those accidents are due to insufficient processing power? All of these make a big difference.

As for alternators, they are practically free, in context. An additional few hundred watts on a belt-driven alternator is extremely minor.

It really isn't free; this isn't debatable. 500W would put a noticeable load on the engine. Keep in mind here we're talking commuter vehicles - nobody is interested in an autonomous sportscar or flatbed truck, so low-displacement 4-cylinder engines are the focus here. The average alternator generates up to 1400W, and that's on the upper end. A 500W computer is consuming 35% of that power. Now, consider driving at night, with the speakers cranked up, charging your phone, headlights on, on a windy road (keep stuff like electric power steering in mind). At this point you are easily exceeding 60% load of your alternator. Like a computer PSU, you generally shouldn't go that high. You will put a lot of wear on the alternator and the drive belt, and your fuel economy will suffer.

A couple hundred watts of headlights on vs off will hardly effect MPG. You don't see people complaining that their MPG sucks at night or when they turn on their stereo system do you? It's because even combined, the difference is negligible.

That's because you're well below the limits. Tacking 500W on top of all of that will make a difference. An alternator's efficiency is not linear.

and lastly, if you compare TOPS performance, you will see that the 500w system is many times more energy efficient(perf/watt)

I agree that the hardware is very efficient for it's capabilities. My point is the approach to their system is inefficient.

schmidtbag:

There are multiple versions; the Tesla version uses less power. And even then, 250W isn't that bad. Its still high but not "this may need liquid cooling" high. As for the capability, you're getting into a bit of a gray area here because that tends to be more up to the software rather than the hardware. As far as I'm concerned, Nvidia is just accomplishing their goal without any optimizations. I'm sure it's handling more sensor data than it needs, I'm sure the code could be written more efficiently, and as a result, they don't need that much processing power. I'm not faulting them for that specifically; as far as I'm aware, they haven't been doing this for as long as other organizations, so they're making good progress.

Tesla has replaced the hardware on it's vehicle several times already as they move the goal posts for FSD and they are doing it again with custom designed in-house hardware. So until we know the power consumption on that we have no idea what the real compute/power requirement for FSD is. As far as software optimization they're never going to train a deep learning system to understand the contextual clues a human brain can (at least not in the next decade or two). They supplement that lack of "understanding" with more sensor data to avoid accidents and like you said to avoid oddities with the tech and reflective surfaces, etc. You have to remember a lot of this started out with lots of people making really bold claims - Elon claimed you didn't need radar or anything, you could just do it with vision and now that's out the window. George Hotz claimed he could do it with just a cellphone.. lol. As far as power requirements, I think it's a non-issue. It's a single generation of hardware, the following generation will probably have half the power for similar performance.. if that's truly the performance requirement for FSD. They can redesign the power delivery systems to handle the load. Plus I doubt anyone is going to care about 10-15 mpg loss when they don't have to drive anymore, I personally wouldn't, I'll gladly pay that extra fuel/power cost to just sit back and watch a movie on my daily ~1Hr commute through NJ garbo traffic. Long term FSD just completely changes everything in terms of efficiency. No one will own cars anymore, you'll just sub to a monthly fee and when you want to go somewhere just summon one and go. The vast majority of cars spend something like 80%+ of their life in a parking lot. So you're going to yank millions of cars off the road. Plus can eliminate tons of parking - which will make city parking/design easier and more efficient. Traffic Elimination because Grandma Georgina is no longer going to merge into a 65mph highway at 30. Also Nvidia has been testing FSD cars out of their Holmdel NJ office since at least 2013. I've seen them on the garden state parkway several times. I don't think they are as far along as Waymo is but I think they are further along than Uber/Tesla. The difference is that Nvidia is shipping this as a system to car manufacturers - so there is most likely a significantly larger liability there. They need to be 10,000,000% sure the system is safe before they start activating it across Volvo/Audi/Etc.

Denial:

Tesla has replaced the hardware on it's vehicle several times already as they move the goal posts for FSD and they are doing it again with custom designed in-house hardware. So until we know the power consumption on that we have no idea what the real compute/power requirement for FSD is. As far as software optimization they're never going to train a deep learning system to understand the contextual clues a human brain can (at least not in the next decade or two). They supplement that lack of "understanding" with more sensor data to avoid accidents and like you said to avoid oddities with the tech and reflective surfaces, etc. You have to remember a lot of this started out with lots of people making really bold claims - Elon claimed you didn't need radar or anything, you could just do it with vision and now that's out the window. George Hotz claimed he could do it with just a cellphone.. lol.

I think it's reasonable to assume that they would try keeping the power envelope pretty similar to what they had before, since they can't really afford to increase it a whole lot without people noticing an impact on range. As for the deep learning thing, I don't understand what you're getting at. It'd be a terrible idea to expect each individual driver to train their car's AI, so that obviously isn't how they would go about it. That means the AI is something they (as the company) would train, and only need to do once. I'm sure they'd release small updates every once in a while to fine-tune its performance. It's so much less expensive in terms of hardware, energy, financial cost, and processing power to have a trained AI, vs the alternative you propose (compensate with a crapload of sensors). It's something that, if done properly, would apply to everyone's car. To clarify my point in my previous post, sensors should only be added to compensate for missing and/or unreliable data. But just simply adding sensors for the sake of more data will not necessarily ensure a smarter or more effective AI. And no, I'm not suggesting it'll hurt the AI (though, it could, if there's enough latency or incoming errors as a result of them).

As far as power requirements, I think it's a non-issue. It's a single generation of hardware, the following generation will probably have half the power for similar performance.. if that's truly the performance requirement for FSD. They can redesign the power delivery systems to handle the load. Plus I doubt anyone is going to care about 10-15 mpg loss when they don't have to drive anymore, I personally wouldn't, I'll gladly pay that extra fuel/power cost to just sit back and watch a movie on my daily ~1Hr commute through NJ garbo traffic. Long term FSD just completely changes everything in terms of efficiency. No one will own cars anymore, you'll just sub to a monthly fee and when you want to go somewhere just summon one and go. The vast majority of cars spend something like 80%+ of their life in a parking lot. You can eliminate most parking lots as well - which will make city parking/design easier and more efficient.

500W is not a non-issue, but, I would entirely agree with you that future generations are bound to be far more efficient. However, Nvidia are the ones claiming that 500W is good; it isn't. As stated before, it's efficient in a TOPS perspective, but performance efficiency is not the only metric that matters. If it did, everyone would be driving 4.0L+ engines, because despite using more fuel, a larger displacement is more efficient (when all things are treated equal - obviously, plenty of engines that size are horribly inefficient). EDIT: Also, it's important to consider that many car companies are really being nagged to improve fuel economy, especially in places like CA and various European countries that have emissions taxes.

Also Nvidia has been testing FSD cars out of their Holmdel NJ office since at least 2013. I've seen them on the garden state parkway several times. I don't think they are as far along as Waymo is but I think they are further along than Uber/Tesla. The difference is that Nvidia is shipping this as a system to car manufacturers - so there is most likely a significantly larger liability there. They need to be 10,000,000% sure the system is safe before they start activating it across Volvo/Audi/Etc.

Interesting, I wasn't aware of that. Well then, I guess their system is better than I gave them credit for.