In early August, Tesla CEO, Elon Musk, announced that his company had developed a new AI chip for his Electric cars and claimed that it is ten times faster than the ones they use today from nVidia. When I was at NVIDIA’s spring developers conference, I went to a session on NVIDIA’S AI Xavier processor that is targeted for use in cars. In this session, I learned that at that time, it was the most powerful processor for AI available. When Musk made this statement about his own processor, I assumed he was comparing it to Xavier, and if it were ten times faster than this chip, he would have achieved quite a feat.
But a call to Danny Shapiro, NVIDIA’s Director of Automotive Technology, spread some light on Musk’s claim.
“The performance claims are against what they have in the vehicle today, which are three years old,” says NVIDIA’s director of automotive Danny Shapiro. NVIDIA’s latest silicon is at least ten times faster than that, which would put it on a par with Tesla’s chip.”
NVIDIA has been a critical partner for Tesla, and if it were not for them, Tesla cars would not be on the market today. What I disliked about Musk’s proclamation about his own processor is that he did not clarify that he was comparing it to an older chip, and in essence, threw NVIDIA under the bus. That is not the way to treat a partner that bent over backward to work with Tesla and help them achieve the success they have today.
That said, Tesla is moving towards vertical integration by creating their own AI chips and, where possible, producing components for their cars internally. In fact, they make their own seats for their cars already, and I suspect that over time they will try and move more and more towards creating as many components for their vehicles they can to control the supply chain themselves.
Google, Facebook, and others are doing something similar. The # 1 thing they are trying to take control of is the core CPU and GPU needed for any ancillary hardware products they make to extend their products and services. In Google’s case, they are creating chips for their smartphones, laptops and AR and VR headsets.
In Facebook’s case, they are creating AI chips for use in their Oculus VR headsets and who knows what other actual hardware they may build over time. The idea of taking control of a company’s eco-system of various components are not new. Indeed, Apple is the poster child for this type of vertical integration.
Early on, Apple obtained a core license to the Arm architecture and had built their own custom-semiconductor on top of an Arm core. This has allowed them to control their own destiny at the chip level. To date, Apple still uses Intel chips in their Mac desktop and laptop products, but I would not be surprised if Apple, in the next two-to-three years, migrates the Mac OS over to work with the Arm-based processors too.
At the heart of these moves towards vertical integration is the desire to control a company’s destiny. Apple has been criticized for the slow updates they deliver to the desktop and laptops, but that is mostly due to their reliance on Intel’s processors and when new ones become commercially available.
It is also now much more comfortable for a company to create their processors by either doing what is called semi-custom design, which is what they do when they build on an ARM core for example. In more extreme cases, they might develop the chip from scratch. In the past, for them to develop these chips, most had to create their own fabs to make them. Today, thanks to fabs like Global Foundries, TSMC, Samsung, etc, they can take their design to these semiconductor manufacturers and let them build it for them.
However, there is a risk when doing custom Silicon. In a recent Tech.pinions subscriber post, my son Ben, who is much more semiconductor literate than I am since his background is in chips, wrote about what is at stake when one’s does custom Silicon:
By going down the proprietary silicon route, companies are betting their differentiation on this strategy. I say it this way because of one common theme with companies efforts in custom silicon is how focused it is on a specific area of differentiation. If one did an in-depth analysis competitively, you would likely find where these companies believe they are differentiated by where they try to build custom silicon to enhance or deepen that differentiation.
The risk, however, in this scenario is their efforts in custom silicon are not able to sustain their differentiation as companies like NVIDIA, Qualcomm, Intel, AMD, etc., design better, more powerful solutions that enable said companies competitors to have an even better offering.
While I agree with Ben about the risks, I suspect that these companies will plow ahead with their own custom silicon solutions anyway. Their desire to control their own destiny even at the chip level and use it to differentiate seems too critical to their future for them to trust any given supplier to meet their futuristic vision with off the shelf processors.