Tesla Is Working With AMD To Develop Its Own AI Chip For Self-Driving Cars (cnbc.com)

An anonymous reader quotes a report from CNBC: Tesla is getting closer to having its own chip for handling autonomous driving tasks in its cars. The carmaker has received back samples of the first implementation of its processor and is now running tests on it, said a source familiar with the matter. The effort to build its own chip is in line with Tesla's push to be vertically integrated and decrease reliance on other companies. But Tesla isn't completely going it alone in chip development, according to the source, and will build on top of AMD intellectual property. On Wednesday Sanjay Jha, CEO of AMD spin-off GlobalFoundries, said at the company's technology conference in Santa Clara, California, that the company is working directly with Tesla. GlobalFoundries, which fabricates chips, has a wafer supply agreement in place with AMD through 2020. Tesla's silicon project is bounding ahead under the leadership of longtime chip architect Jim Keller, the head of Autopilot hardware and software since the departure of Apple veteran Chris Lattner in June. Keller, 57, joined Tesla in early 2016 following two stints at AMD and one at Apple. Keller arrived at Apple in 2008 through its acquisition of Palo Alto Semiconductor and was the designer of Apple's A4 and A5 iPhone chips, among other things. More than 50 people are working on the initiative under Keller, the source said. Tesla has brought on several AMD veterans after hiring Keller, including director Ganesh Venkataramanan, principal hardware engineer Bill McGee and system circuit design lead Dan Bailey.

Comments

Incorrect story

[TimothyHollins]

According to Tom's Hardware, this story is a misunderstanding, and does not represent the actual words of the presentation.

From TH:

Some media outlets are reporting that GlobalFoundries is working with Tesla on AI technology for its cars. This erroneous report stems from a comment GloFlo's CEO Sanjay Jha made on stage on Wednesday at the fab's annual get-together in San Jose. ...

But what Jha actually said—which we can confirm because we were present to hear it firsthand—was that GlobalFoundries is trying to attract companies as business models change:

"As we develop these new technologies, we are also seeing a big shift in the business model and the foundry business. What is happening is that system companies like Google, like Amazon, like Tesla, like Microsoft, are coming directly to foundries. They are working directly with IP companies and system development companies because they want to control the hardware and software."

Global Foundries is not saying that it's working with Tesla--but that's not to say that AMD isn't working with Tesla. Jim Keller, formerly the chief architect for AMD's microprocessors, is now VP of autopilot hardware at Tesla.
Last year, AMD lost what Tesla CEO Elon Musk called a tight race against Nvidia for the auto company's GPU/AI business. Since that time, AMD has continued to show strength across multiple sectors.
The CNBC report said that its sources tied AMD and Tesla together, but neither AMD or Tesla will comment on the situation. The report indicated that Tesla was on a mission to develop its own chip for autonomous cars in order to be more vertically integrated, but that Tesla was potentially relying on building that "on top of AMD intellectual property." That particular wording certainly paints a dotted line to GlobalFoundries.

Full story at http://www.tomshardware.com/ne...

Auto companies, patents, etc

[david.emery]

Florian Mueller predicts the (German) auto companies will become patent trolls, as the tech industry takes over autonomous car design:
http://www.fosspatents.com/201...

Are we going to see a convergence, where tech companies and auto companies team up, or a divergence, where tech companies produce the new vehicles and legacy car companies shrink into irrelevance?

The only thing I can predict with great confidence is that the cost for a replacement CPU board for a Tesla will be A Lot More than the cost of the constituent parts. (Nissan charged me $1500 for a truck wiring harness after mice chewed the insulation. It's really hard to believe that almost 6% of the cost of that truck was in the wiring harness.)

Re:Auto companies, patents, etc

[0100010001010011]

The tech companies are going to be in for an entertaining world of hurt if they think they can just show up and start cranking out products the way they've always done it.

Automotive, industrial, aerospace, rail and other 'life and limb' companies have done things their own way for a while for a reason. We have functional safety standards that have no 'tech company' equivalent (that I've seen). ISO 26262, IEC 61508, DO-178C, ASIL A-D, etc. From what I've found Intel and AMD don't have any chips that meet those standards.

Your phone crashes you file a bug report. Grandma crashes and dies and it's not just 'file a bug report'. The German automakers have been working on all of this stuff for a while. In grad school years ago we had some researchers from VW come in and show off all of their collision avoidance designs. I rented cars in Germany that had the auto-stop feature that didn't show up in the states until years later.

They're not just patenting stuff, they're implementing it. The difference between them and Tesla is they're very reserved in what they say it can and can't do and don't call their shots before they make them. Tesla drives up hype says what they will do and hopes they can deliver.

You think Cummins and Benz heard about the Tesla truck announcement and threw together a full prototype in a month to beat them to the punch? People are naive if they don't think that everyone is working on the exact same self driving features Tesla is. They're just not as vocal about it.

If you want to know what those companies are and when they got their start, look for the logos plastered all over the DARPA 2004/2005 vehicles. It's not like they grabbed their $1M, split it up, patted themselves on the back and forgot about the technology. All of those grad students got hired on. Startups that formed out of it got bought up by bigger players. We've all been working on this stuff, together, behind the scenes getting it right and testing it off line before we brag to the local newspaper.

Wtf are you talking about?

[Viol8]

"given today's technology and established code base, for them to use anything but inexpensive, commodity parts for this"

This is a highly specialised application. Sometimes commodity parts are too inefficient and/or slow to be a good choice in these situations.

"there are a lot of people who can code for them"

With modern compilers being able to code for a specific chipset is less relevant. Also while there are many experienced to-the-metal assembly language coders and there are many experienced people who can code a neural net or other AI paradigm, I suspect the intersect between the two is quite small.

" I think this is simply a culture issue with Tesla wanting to prove that they don' need no one else, they're just THAT good, nanny nanny boo-boo"

No, I suspect its a business decision. And it makes a chance having a corporation who wants to build the best and not just shove out cheap consumer level shit that goes wrong after a couple of years.

Oh, and you might want to ease up on the schoolboy mocked, it doesn't add weight to your argument, it just makes you sound like an ass.

"roll their own outside of their areas of core expertise."

I think this whole discussion is certainly outside your core area of expertise.

Re:Wtf are you talking about?

[Rei]

It's simpler than that - if Tesla adopts it, it will become a commodity part. Tesla is looking to start Model 3 full produciton at 500k per year, scaling up eventually to 700k per year, and then do the same with the Model Y, all just in the next several years (not to mention future products). Whether or not they manage to achieve their goals, these are their goals. Every last one of them will come with Autopilot hardware, regardless of whether or not the person purchased Autopilot. Millions of units per year is pretty much "commodity hardware" in my book.

It makes a lot more sense to have specialized hardware that can run the task much faster with much less power than to use hardware not designed for the task when you're dealing in these sorts of quantities.

Complete BS

[Groo Wanderer]

I was at the Global Foundries event and the keynote, no such thing was said. The Keynote recordings did not say that either, Tesla was mentioned as an example but the article is badly off base, so badly that it seems intentional. I checked with the speakers in question, other journalists, and the PR people at the show, ALL confirmed the story was not true and what was claimed to have been said was not.

            -Charlie

Sounds like marketing bullshit

[Rick Schumann]

This really sounds like more marketing/media hype that doesn't know what the hell it's talking about, even graded on the scale of the usual clueless, inaccurate so-called 'AI' marketing/media hype. How the hell do you have an 'AI chip' when we don't have anything like actual 'AI' to start with? Also how the hell do you have a 'self driving car chip' when the technology realisitically isn't even really close to ready for the general public? I think the 'translation' of this 'story' is this: They're getting some proprietary version of an otherwise garden-variety CPU or SoC; in other words, 'Nothing to see here'.

Re:Sounds like marketing bullshit

[Rei]

How the hell do you have an 'AI chip' when we don't have anything like actual 'AI' to start with?

Nobody is talking about "generalized AI". The topic of concern is very specialized AI tasks. We have a wide variety of AI systems in use for specialized tasks in our everyday lives - particularly in the topic of image recognition.

Also how the hell do you have a 'self driving car chip' when the technology realisitically isn't even really close to ready for the general public?

It's already very popular among many Tesla users (particularly if you're stuck in traffic). But nobody is talking about the entire software stack being implemented in hardware and thus being permanently locked down. What's being discussed is offloading specific CPU-intensive image and radar processing tasks to the chip.

Re:Wrong approach

[AaronW]

For AI/machine learning the parts are rather specialized. For example, look at Google's tensor processing unit (TPU). Machine learning typically involves a lot of large integer matrix multiply operations where dedicated hardware makes a big difference in performance.

For example, Google's TPU consists of 64K 8-bit multipliers in a 256x256 array along with around 4M adders. A general purpose CPU or even a GPU will not be nearly as optimal. Google's TPU performance per watt is around 83x as good as a CPU and 28x as good as a GPU.