On Tesla’s Incredible Platform Advantage
Since the Model S was released anyone who has ridden in the car says it feels like magic. The model 3 extended on this, arriving late but suggesting Tesla – as a car company – honed an ability to produce experiences far above most others, even though they started from scratch.
There are reasons why riding in a Tesla feels better than other cars most people have ridden in, and it’s not entirely clear whether it’s the subtleties, the raw speed or what’s under the hood that plays the larger part. When you speak to owners they say a Tesla presents itself in a way that lets you feel unconsciously in control yet expressive and free.
Yesterday we saw news from Tesla that showed they have been working on their own chips, basically a deeply integrated SoC they call the Tesla Supercomputer, built using its own design team and optimized for Samsung’s fab, their own approach to comparable technology from Nvidia. Many people focused on the chip’s performance, when the interesting part is actually Tesla’s emerging platform advantage and the strategic ramifications of switching from external suppliers.
It always surprises me how few people talk of or understand vertical integration at the deepest levels of what originally made Silicon Valley—the silicon.
One of Steve Jobs’ biggest legacies was his decision to stop relying on 3rd party semiconductor companies and create an internal silicon design team. One of the things that was common about Apple’s original approaches and Tesla’s effort now, is the former PA Semi heritage of the team. But there are other interesting parallels too.
The reason this new chip exists in 14nm instead of 10nm process technology is complex. In order to move fast, Tesla licensed IP blocks — things like high speed analog and SRAM libraries — which allows it to focus on the key parts of the IP — the neural network. This strategy is common in complex SoCs. When Apple designed its chips originally, they couldn’t tackle all elements of the A-series and used a little known company called Intrinsity for some of the IP blocks, eventually buying them and folding IP into the team.
The key to moving fast for Tesla is making complex trade-offs between backward compatibility and future optionality, which will be derived on the fly in software. When analysts asked about why the chip is fabbed in 14nm vs 10nm, Tesla’s head of architecture answered flatly that some of its licensed IP libraries were only available in the older 14nm process tech.
The challenges of spinning up a design team and building vs buying IP are multifold, it’s about things like compatibility and IP, budget, cost, speed and risk.
The key to understanding What Tesla did in releasing what they call a supercomputer is understanding the 2019 version of building and supporting a full custom semiconductor design team. The mere fact that they are already shipping this computer retrofitted to an existing form factor in Model X, S and 3 is astounding. This was not a preannounce by Tesla. This was an in-production launch.
And the key to understanding Elon Musk is to separate what is hyperbole from fact—he uses both skillfully. The most important truth from watching yesterday’s unveiling is that this chip is not only in full production from fabs, but also shipping already inside cars.
Saying you are going to do something like full autonomy on ‘X’ time-scale is one type of commitment, but saying you built an entirely new rendering engine for the pixels that fly by in the physical world is quite another. While any former projected launch for ‘full autonomy’ is supposition, marketing, and subject to industry definitions, demand and regulations, the latter is something Tesla can — and did — control.
In 2015, when people famously declared there was no chance the model 3 would arrive in time and reach its cost and market share goals, Elon was off creating a chip design team. If you study unit economics of semiconductors, it doesn’t really make sense to design chips and compete with companies like Nvidia unless you can make it up in volume and differentiate considerably from application specific chips which are mass market. Consider the audacity back in 2015 for Tesla to believe it could pull this off. How would they ever make back the R&D to build out a team and pay for expensive silicon designs over the long run, never mind design comparative performing chips?
It is – in fact – chip making capabilities, that Jobs famously brought in-house to Apple shortly after the launch of the original iPhone, that helped Apple create a massive profit moat between itself and an entire industry (smartphones).
Chip advantages ultimately became one of the little understood, but critical elements in Apple’s vertically integrated approach. Android OEMs can copy the Face ID sensors. They just go to the camera supplier that Apple buys it from. But they can’t copy the underlying software powering and securing its own chips.
One lens that helps assess the asset value of a buy vs build strategy is called buyer / supplier bargaining power. Because of Tesla’s small scale in cars they don’t have massive influence with suppliers. What Tesla gets in chip land is what everyone else gets.
Since semiconductor design cycles are typically 2-3 years minimum for a full custom chip like what Tesla built, buying this amount of time is really valuable—other car companies must react on much shorter scales to what is available in market and can’t plan ahead by integrating and owning software at the neural engine and frequency / image processing levels.
The challenges in owning a chip team today in the Valley are immense. It’s important to realize these designers who just released this chip in to real partially autonomous cars are motivated now… make no mistake. Their ‘computer’ is in the wild. Tesla will be getting immediate feedback from real cars. Even if Tesla only ships a few hundred thousand vehicles for the remainder of the year, this feedback is real and informs the next design.
But vertically integrating at this level does much more as Tesla aims to improve autonomy. Making chips could serve as a moat around untold strategic advantages: development secrecy, hyper optimization of image processing algorithms, data security, predictive AI, etc. And all the while their modest yet real car volume serves as ‘R&D lead gen’ for the new chip and software which will power Tesla in 2 years. By owning its own silicon design team, Elon is able to leap into new autonomy problems which will be eaten by software running on its own silicon.
Building competent semiconductor design capabilities is an absolutely massive endeavor. Despite its original moniker, the Silicon Valley produces almost no venture backed chip companies today. Instead, talented designers post up at the existing old-guard companies such as Qualcomm and Intel. What chip designers crave is seeing their designs solve new and novel problems. Creating and staffing a top caliber team is not for the faint of heart. The fact that Tesla embarked on this and funded it through its dark days of Model 3 hell is unprecedented.
Because chips in cars are relatively low volume, no other car maker will be able to enter this game. It’s simply not in their DNA, and volume would also never allow it – sure other car makers have a lot more volume than Tesla, but it’s still two orders of magnitude below smartphones. Both the costs and the risks of designing chips are way too high. Tesla doing this from the Silicon Valley itself is both a by-product of its strategy of “being there” as well as its ability to recruit top designers and architects who buy into the dream and want to be challenged.
But when you look through the lens of the car as a mobile device, the software and silicon look markedly similar. It’s clear autonomous cars will rely on application specific chips and off the shelf sensors to accomplish the feat; it’s just not clear whose chips and which sensors. By building a massive neural engine capable of step function increases in safety / reliability over basic human consciousness, Tesla can push vision processing to its absolute and utter limits. This alone means that the vision approach vs LiDAR may have real legs. Sure, what Elon said about LiDAR being the losing technology is hyperbole, but don’t forget that what he said today about their new chip is not.
The question is not really when Level 5 autonomy will be on the road—it’s who can survive long enough to get there.
One posit you can derive from this is that no other car manufacturer in the world can support an internal chip team. The low volume nature of cars and the out-of-Silicon-Valley predicament preclude anyone being audacious enough to even attempt to pull this off, which makes Tesla’s feat feel much more pronounced when you rationalize this.
How does this impact fully autonomous cars? A Tesla can be exclusively Tesla under the hood—aside from the image gathering sensors themselves. Others will be effectively running off the curve both of LiDAR vs image processing as well as the OS and chip level frameworks supplied by Nvidia and anyone else they need to depend on in the standard supply market. For Tesla this is all about custom software at the bleeding edge and strategic learning.
Just as there are many misunderstood reasons for Tesla’s overall success, justification as to how Tesla produces superior cars isn’t superficially visible. This is always the case with companies who produce and sell magical products or experiences.
But one thing is very clear: we are entering an era where cars will become autonomous, navigate by themselves and prevent catastrophes from happening at unprecedented scale. And just as smartphones spawned a different set of winners every decade, you can bet the next set of car makers will look much different than today’s. Google, Uber and Tesla will all be involved… And although we know almost nothing of Tesla’s future chip plans, we know they did something uncharacteristic and deeply impressive. They built a hardware and software platform foundation from the ground up, funding and subsidizing it – just barely – based on the incredible dream they are selling today.