When you click on links to various merchants on this site and make a purchase, this can result in this site earning a commission. Affiliate programs and affiliations include, but are not limited to, the eBay Partner Network.
it's not different than say apple which does its own software and cpu design but apple does of course OUTSOURCE almost all of its manufacturing so complete vertical integration isn't necessary at least in their case..
No different on the surface but vastly different if we look at the details. Tesla made their own CPU or GPU or whatever they call it with no experience, and never having made one before. And being a new company. Apple made their own after they amassed vast resources and experience.
On vertical integration this is in theory always better but you have to do it right. Something I don't see talked about is the possibility many OEMs didn't want to work with Tesla, didn't want to upset their long held relationships. Plus none of them can make the core components for Tesla anyway they don't know how.
No different on the surface but vastly different if we look at the details. Tesla made their own CPU or GPU or whatever they call it with no experience, and never having made one before. And being a new company. Apple made their own after they amassed vast resources and experience.
Lol, CPU or GPU or whatever you call it. Do you expect your arguments to be taken seriously?
Lol, CPU or GPU or whatever you call it. Do you expect your arguments to be taken seriously?
Here's a die shot
See the section that says GPU, and the section that says NPU (this is a fancy name for a CPU). You were saying? Now since you know all about it what does Tesla call the chip specifically could be called SoC, not that it really matters.
adding 'npu' labels on the die is tech b.s. to make it seem like it's smarter than it is (like auto pilot that is neither auto nor pilot). no doubt tesla has some serious custom processing power in there but i don't believe for a second that other can't catch up. plus the possibility of ACTUAL self-driving tech firms like Google (Waymo) licensing their tech to car makers.
Would you trust Google to drive your car? I wouldn't. Plus you're guaranteed they will sell your telemetry. When do you think Google will catch up to Tesla? How many cars use Waymo?
ACTUAL self-driving tech firms like Google
Actual? Hmmm, Tesla has a real time network and their systems running on what 400,000 cars what does Google have. lol once again some people are hell bent on never giving Tesla credit for anything it has become the biggest whataboutism of all time.
adding 'npu' labels on the die is tech b.s. to make it seem like it's smarter than it is (like auto pilot that is neither auto nor pilot). no doubt tesla has some serious custom processing power in there but i don't believe for a second that other can't catch up. plus the possibility of ACTUAL self-driving tech firms like Google (Waymo) licensing their tech to car makers.
Anyone with even a little bit of common sense realizes that Tesla's "autopilot" is a primitive gimmick compared to Google's system. The sensors Tesla uses can't hold a candle to the LIDAR, the autopilot is only good for highways with proper lane markings, and even then it can't see cross traffic, even if its a giant semi. If the lane marking is worn out and there is no car in front of the Tesla for it to track its path, it will drive right into a concrete barrier. Obviously there are people willing to pay 5k for this option - more power to them, and even more power to Tesla for capitalizing on it.
That being said, Tesla and the rest of the tech companies developing "autonomous driving" are shooting themselves in the foot. Instead of all developing their own systems, both hardware and software, they should be working together on creating a single standard so these vehicles can then properly communicate with each other and traffic devices. Ditto for electric charging infrastructure.
See the section that says GPU, and the section that says NPU (this is a fancy name for a CPU). You were saying? Now since you know all about it what does Tesla call the chip specifically could be called SoC, not that it really matters.
I love beautiful die shots.
Originally Posted by bitkahuna
adding 'npu' labels on the die is tech b.s. to make it seem like it's smarter than it is (like auto pilot that is neither auto nor pilot). no doubt tesla has some serious custom processing power in there but i don't believe for a second that other can't catch up. plus the possibility of ACTUAL self-driving tech firms like Google (Waymo) licensing their tech to car makers.
As a former Intel engineer and having worked on several chipset generations, chip design is incredibly difficult. On the surface it's all transistors, pun intended. The challenge is balancing real estate and making optimum use of those transistor counts to form an architecture that can provide the performance, memory bandwidth, parallelism, feature set, scaling, and stay within a power envelope.
Sure, other companies can do it too, but the number that can is smaller than you think. There are only a few truly genius chip designers responsible for many of the great microarchitectures of the last decade or two -- and Tesla poached two designers that formerly worked at Apple and AMD.
Semiconductor firms compensate their chip architects very well to ensure they don't leave because not too many others can do the job well. This is like the Olympics of engineering. There are many super smart people in all fields but only a few true standouts that can create something world-class from scratch. Not to mention AI/Machine Learning is a burgeoning frontier and legacy automakers (and their suppliers) have not invested to the same degree as a Facebook, Amazon, Microsoft, Google, IBM, or even Tesla. Can other companies match the system design of Tesla? Yes. Are any of those companies automakers? Doubtful at the moment.
Last edited by FatherTo1; Feb 19, 2020 at 01:28 AM.
As a former Intel engineer and having worked on several chipset generations, chip design is incredibly difficult. On the surface it's all transistors, pun intended. The challenge is balancing real estate and making optimum use of those transistor counts to form an architecture that can provide the performance, memory bandwidth, parallelism, feature set, scaling, and stay within a power envelope.
Sure, other companies can do it too, but the number that can is smaller than you think. There are only a few truly genius chip designers responsible for many of the great microarchitectures of the last decade or two -- and Tesla poached two designers that formerly worked at Apple and AMD.
Semiconductor firms compensate their chip architects very well to ensure they don't leave because not too many others can do the job well. This is like the Olympics of engineering. There are many super smart people in all fields but only a few true standouts that can create something world-class from scratch. Not to mention AI/Machine Learning is a burgeoning frontier and legacy automakers (and their suppliers) have not invested to the same degree as a Facebook, Amazon, Microsoft, Google, IBM, or even Tesla. Can other companies match the system design of Tesla? Yes. Are any of those companies automakers? Doubtful at the moment.
You hit the nail on the head here with these observations in bold (see above)
Early on, Musk has had it's computer gaming unit excel in League of Legends and several Esports. Although unrelated to Auto Tech, the concept of AI being applied to real-time gaming is not much different than what a car will face when getting from point A to B. This was several years ago.
Tesla has somehow quietly progressed through 3 iterations of this AI hardware infrastructure for the car unnoticed. Now that the general market is realizing what they just did in recent Model 3 tear down --- everyone else is flat footed and dead in the water on what to do next. Did they make their autonomous capabilities too complicated or too costly or experience latency creep in realtime reaction to a situation? All of a sudden, all automakers are facing the reality that over half a million Tesla vehicles on the road already have the autopilot capabilities in use. Automakers are going to have to figure out real fast what they are going to do about hardware options, software solutions and whether what they are doing going to work or be cost effective for mass adoption. In short, everyone is officially cooked and not ready tech wise because autopilot type of tech can cost as much as the battery/EV. Not everyone will pony up the bucks for it and be a test guinea pig. Cadillac's Super Cruise got some early kudos and only available in a highly optioned configuration. No one is going to buy stuff if it is optional. The Genius Move by Tesla is EVERY vehicle in their line up has autopilot computer and sensors as standard part of the vehicle's DNA (pending software activation by owner).
Exciting stuff positioning Tesla uniquely as a Tech company and on track for a bonkers 2020 decade in small and large sedan vehicles, small and large SUV and more, all with a 'tech dna' buyers can trust.
Just want to share this other video demonstration of what the Full-Self Driving computer "sees" and assesses at hundreds or thousands of times per second. The software is tightly coupled with the hardware to make this all work. I would be amazed if Bosch suddenly revealed a similar system for automakers. Lidar may offer better resolution and abilities in inclement weather but it's also incredibly cost prohibitive and will remain that way for many years. There was another article about mapping underground terrain beneath roads to offer precise, consistent autonomous driving in all types of weather and regardless of fading or missing lane markings. There are several ways to improve autonomous driving but someone's still gotta put it all together in an affordable package with proven hardware and software. It will be interesting how legacy automakers maneuver and adjust to the changing landscape. Many of the automotive CEOs have been blindsided for 7-8 years now.
Here's an example of FSD visualization correctly detecting the color of traffic lights. It doesnt respond to traffic lights yet, but you tealize the progression that has to take place to reach the end goal. The car has to be aware of the environment first. The decision-making comes later.
Last edited by FatherTo1; Feb 19, 2020 at 10:25 AM.
But here's the rub: if this were a guidance system for a factory floor robot, it is easily good enough to be implemented.
But out on the street where there is no control of the environment, if there is ONE mistake out of a BILLION correct ones, a human can be injured or killed. Litigation and million-dollar judgments.
In other words, I tend to think the only place where automation will be implemented will be in a tightly-controlled environment, such as a carefully-monitored interstate highway.
But here's the rub: if this were a guidance system for a factory floor robot, it is easily good enough to be implemented.
But out on the street where there is no control of the environment, if there is ONE mistake out of a BILLION correct ones, a human can be injured or killed. Litigation and million-dollar judgments.
In other words, I tend to think the only place where automation will be implemented will be in a tightly-controlled environment, such as a carefully-monitored interstate highway.
What if a traffic light glitches and results in a crash? What if a computer system on a commercial jet fails? I could name dozens of examples where technology glitches and people are injured or die that doesn't mean we give up and stop using them.
That's awesome work by MB. The question is why isn't it production ready yet since TWENTY THIRTEEN?? Bring it out! At around the 3:45 min mark, it seems like they mapped out a specific route or are relying on a pre-scanned roadway. In either case, I applaud the progress. Why has it taken so long to introduce the tech in an S-Class where they're less worried about keeping cost down? Unless the work is proving to be much harder than automakers realize. I hope the rest of the industry does catch up, but then you already hear about MB delaying the EQC another year to make it more competitive in the U.S. market.
If that's a reason not to trust or buy a Tesla then that's fine. Just like people should stop buying iPhones because the face scan can be circumvented by a photo or the fingerprint scanner can be beaten by some tape. It's not as if humans aren't naive and easy to trick either.
