Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Tesla is building its own AI chips for self-driving cars

techcrunch.com

Post reply on HN
Print view View original

211–220 of 238 posts

Re: Tesla is building its own AI chips for self-driving cars

#211

Post by ethbro »

TeMPOraL wrote: post #190
ethbro wrote: post #174

Earlier quoted context omitted.

If I trained a child to drive, I'd be comfortable letting them drive at 6. I believe I was around that age the first time I drove (handling steering, with assistance on gas / brake). My memory is that the only difficult part was a lack of power steering on hilly terrain. Same age for walking around town unsupervised: the Japanese do and don't seem to lose many. My original point was focused on the technical feasibili…

And my point is that it's not just about experience-years, it's also about the breadth of knowledge. When driving, humans do not use just the experience acquired from looking at things from behind the wheel. They also use their general understanding of physics, of materials, of concepts related to visibility, traction, not to mention expected behaviour of other people and how to read face expressions.

For an average driver, I'd call BS on that. Humans are terrible at learning complicated things they rarely have to recall.

E.g. the disaster that is the first major snowfall, every year, or automatic transmissions, antilock brakes, and traction control systems becoming standard

The human can only reliably be trusted to keep the car between the lines, stop appropriately, and occasionally make turns. Which is something much simpler to compete with!

Re: Tesla is building its own AI chips for self-driving cars

#212

Post by Robotbeat »

dragonwriter wrote: post #187
throwaway648 wrote: post #163

Earlier quoted context omitted.

Estimating distances robustly is very important for navigation. LIDAR is accurate at relevant distances and most importantly very robust against environmental effects. Without LIDAR you can use stereoscopic vision (like human with two eyes) but that is very demanding, not that accurate, and very error prone. Sure, human can do it somewhat well, but eye and brain are extremely complex things to implement (and it still…

> Without LIDAR you can use stereoscopic vision (like human with two eyes) At any but very close distance, don't humans mostly use a combination of lighting cues, a priori knowledge of actual size vs. apparent size, motion parallax, and other flat-image cues instead of stereoscopy? But, yeah, LIDAR cuts through all that, too.

You can use stereoscopic using the full width of the car as the baseline instead of eye distance. Humans have a pupillary distance of about 60mm which is good for about 10 meters (possibly much more: https://jov.arvojournals.org/article.aspx?articleid=2191614). A Model 3 is over 6 feet in width, so the pupillary distance is thus 30 times that of humans, and so should be good to about 300 meters, which is comparable to high end LIDARs (although the stereoscopic approach won't be as precise at those distances).

Anyway, if LIDAR becomes small and cheap, Tesla can just strap it on.

Re: Tesla is building its own AI chips for self-driving cars

#214

Post by Robotbeat »

twtw wrote: post #26

The most interesting question about this is what their motivation was to justify building an in-house ASIC design team. That isn't a super hard thing to do for a company of Tesla's size, but it is nontrivial and seems like a significant investment. It is probable that they got massive performance improvements by specializing their hardware, but it isn't clear that they couldn't have gotten the same perf improvement b…

It has become easier to make ASICs in-house than it used to be, particularly if you're trying to do something relatively straightforward (versus trying to build a fully featured, backwards-compatible x86-based cpu architecture). The teams which developed Bitcoin mining ASICs were pretty small, for instance (although that's an extreme case of simplicity). You don't need your own foundry.

Re: Tesla is building its own AI chips for self-driving cars

#215

Post by gruturo »

ckastner wrote: post #205
modeless wrote: post #141

Earlier quoted context omitted.

Yes, Musk said that on the earnings call. But the hardware upgrade is not the hard part of self driving. At least they have stopped claiming that their current hardware is powerful enough for full self driving already, which was a ridiculous and worrying position for them to take.

> At least they have stopped claiming that their current hardware is powerful enough for full self driving already The Autopilot page [1] still states the following: "All Tesla vehicles produced in our factory, including Model 3, have the hardware needed for full self-driving capability at a safety level substantially greater than that of a human driver." [1] https://www.tesla.com/autopilot

If you consider “hardware” as sensor suite only (quite a stretch of an interpretation, I know), then it’s kind of true, especially if they will replace the brains at no cost once they have an actual use for them

Re: Tesla is building its own AI chips for self-driving cars

#216

Post by baybal2 »

pasta wrote: post #33
gallerdude wrote: post #6

The benefit they point out is 200fps vs. 2000fps, but is that actually that useful? In the post of OpenAO's robot, they said that improving the response times gave no significant gains. Only thing I can imagine would be allowances for wider and deeper neural nets.

Let's say you drive 130km/h (36m/s). 36 / 200 frames = 18cm travel distance between each calculation. So you either can calculate much more each 18cm or calculate shorter distances. I think a shorter distance is not very useful but calculating more might be.

One thing high FPS are good for are optical flow mappers. Think of estimating relative motion of things in video.

Re: Tesla is building its own AI chips for self-driving cars

#217

Post by baybal2 »

xedeon wrote: post #135
DmenshunlAnlsis wrote: post #5

Earlier quoted context omitted.

Just wait until they admit that they need LIDAR too... Cha Ching! Various Autopilot accidents strongly imply that whatever the computer “brain” is doing doesn’t matter if it doesn’t have sufficient “eyes” to actually see things like a fire truck, cop car, or bollard.

Again with the LIDAR argument... Humans drive using only eyes (computer vision) and having a sense of location (localization/3D maps). Early autonomous test vehicles used LIDARS because camera and compute tech are not at the same level they are today. It's a legacy system and more of a shortcut. In essence, Object recognition > LIDAR judging distances. Yeah it's true that a vision based system is much harder to solve…

Plain radars are still vastly superior to all and every vision system on the market, and don't require you to carry around a supercomputer.

Even if they "crack" it, they will just be spending much more money to get the same benefits as a $100 mm wave system.

In the future, vision systems will only be seen as a gimmick of this era.

Radars beat both ladars and vision systems at phase, speed, and distance estimation.

Re: Tesla is building its own AI chips for self-driving cars

#218

Post by GreenPlastic »

Fricken wrote: post #45

Right now, with HW2 Tesla is in kind of an awkward no man's land where their hardware is overkill for basic ADAS, but nowhere near enough for any kind of autonomy that allows the driver to turn their attention away from the driving task. By going to a custom ASIC with inferencing. I have little doubt about Musk's claim that it'll offer a 10x improvement over the current stripped down PX Drive2 variant for deep learni…

Are they hardware limited? It seems like the limiting factor is the convnet rather than the hardware...do they think it would be more safe with less latency?

Re: Tesla is building its own AI chips for self-driving cars

#219

Post by cma »

antris wrote: post #113
cma wrote: post #103

Earlier quoted context omitted.

Other people are going to be using inferencing chips soon too. Tesla might even just be basing it all on NVDLA (Nvidia's open source inferencing hardware targeted at self driving) and calling it their own.

Of course the competitors will follow. But Tesla will probably be first and a step ahead in the future, hence the talk about go-to-market speed.

Waymo has possibly already been using it for years as Google pioneered that type of chip.

Re: Tesla is building its own AI chips for self-driving cars

#220

Post by mhh__ »

IMTDb wrote: post #162
titzer wrote: post #65

Earlier quoted context omitted.

Humans are not remotely programmable.

Ever heard of the concept of "propaganda" ?

There is obviously a difference between hacking a single vehicle and (say) convincing a generation of youth to die for a cause.

Also, while I agree that you can "program" a human, you can't program them to jump up and down and touch their toes (so to speak)

Post reply on HN