Samsung Potentially Takes the Lead on Tesla’s Self-driving Chips
How will Samsung's involvement affect Tesla's silicon-level trajectory in autonomous driving?
Following earlier rumors of TSMC’s involvement, new reports have pegged Samsung as Tesla’s supplier for next-generation self-driving chips. Industry sources claim that the duo is jointly designing and fabricating Tesla’s HW 4.0 chip, which will enable advanced self-driving capabilities in upcoming vehicles.
However, while there’s a strong belief that Samsung shall triumph, neither company has given confirmation.
Samsung is reported to be heading production for the HW 4.0 chip. Image used courtesy of Samsung and The Korea Economic Daily
Who Will Take the Lead on HW 4.0 Production?
It’s speculated that Tesla’s prospective HW 4.0 semiconductor will be based on Samsung’s 7nm process—instead of the company’s latest 5nm alternative. With shortages still looming and production hikes expected, yields have become a driving force in the manufacturing space.
Because Samsung’s older, larger process’ production kinks have been ironed out, it’s allegedly been dubbed the “safe” choice for the HW 4.0. This “brain” of the car will also be called the Full Self-Driving (FSD) Computer 2.
Samsung, however, does have one thing going for it: it already produces Tesla’s HW 3.0 chip. This predecessor to the 4.0 already powers a number of vehicles in Tesla’s lineup. Samsung and Tesla have a history together, and there’s assuredly some level of comfort in continuing that partnership. KED Global’s report explicitly mentioned long-term cooperation as a contributing factor, alongside affordable production costs and design technology availability.
Samsung already produces Tesla's Full Self-Driving Chip. Image used courtesy of The Autopilot Review
Interestingly, Samsung has also planned to soon build a partially-subsidized U.S. chip plant—dramatically shortening Tesla’s supply chain, should some production capacity favor the HW 4.0.
Overall, optimism over Tesla’s 2022 Cybertruck, and other vehicles, has helped ignite this speculation. That said, what concrete details do we already know about Tesla’s operations?
Tesla’s Semiconductor History
Since 2016, Tesla has poured resources into forming its own, in-house R&D team for chipmaking. Associated silicon-design efforts have spawned a number of lynchpin technologies within the Tesla vehicle lineup:
- The FSD: Powering Tesla’s self-driving software via AI inference, this chip is built upon Samsung’s 14nm process and incorporates three, quad-core Cortex-A72 clusters.
- Project Dojo: Dojo is Tesla’s own AI-training supercomputer, which helps forge the bonds between self-driving neural networks and embedded silicon.
- The D1: This custom AI chip birthed from Project Dojo helps train Tesla’s algorithmic Autopilot driving system.
The FSD computer. Screenshot used courtesy of Tesla
The Full Self-Driving Chip (2019) was Tesla’s first foray into self-developed semiconductors. It was first placed within the Model S and Model X, prior to its inclusion within the Model 3. The original FSD is built upon a 260 mm2 silicon body while packing in six billion transistors. The entire package includes an integrated neural network processor, GPU, main processor, safety system, and security system. This Samsung SoC is a major upgrade from Tesla’s previous NVIDIA chips—offering 21 times greater performance in a smaller footprint.
The FSD has since offered cost savings without compromising on functionality. It is 20 percent less expensive than its predecessor and also draws roughly 21 percent less wattage during use. Accordingly, the FSD has allowed Tesla’s cars to maintain their ranges without added expense.
Understanding the D1
Meanwhile, the automaker’s D1 chip (2021) represents yet another advancement in AI chip design. Tesla’s self-driving philosophy is unabashedly unique. CEO Elon Musk has openly criticized the use of LiDAR systems in autonomous vehicles.
The company instead employs onboard neural networks, which process information from radar units and cameras. These algorithms are integral to constantly understanding the car’s surroundings—in real-time—due to the sheer quantity of data captured by each Tesla vehicle. This indirect scenery mapping requires a massive amount of processing power.
Additionally, the algorithms behind these networks aren’t immediately effective. It takes hours of exposure to label contextual data, called “training sets,” to help them uncover patterns. It’s the process that allows a Tesla to determine what’s a tree, what’s a pedestrian, and what traffic signals must be heeded.
Ganesh Venkataramanan, head of Project Dojo, presenting the D1 chip at Tesla's AI Day 2021. Screenshot used courtesy of Tesla
The D1 chip rapidly accelerates these learning processes. The processor packs 50 billion transistors and delivers 362 teraflops of power. D1 is considered an application-specific integrated circuit (ASIC) and is built upon a 7nm node. However, the chip measures a massive 645 mm2. It bests NVIDIA's top offerings based on processing power.
Interestingly, Tesla has teamed up with TSMC in manufacturing the chip. Should Tesla recruit Samsung for the HW 4.0, as previously mentioned, this would create a notably fragmented production picture.
Overall, the D1 was designed to slash latency and boost performance. As the latest evolution of Tesla’s own technology, it holds a crucial role in enabling next-generation driving systems.
No Formal Agreement With Samsung Yet
Tesla’s designs are meant to outpace any current computational demands. The company has already bested longstanding market leaders in performance while working toward greater efficiency and power savings.
On the partnership front, a concrete agreement has yet to materialize at the time of writing. There are signs suggesting that Samsung has an edge, though recent dealings with TSMC could throw a wrench into those plans. Tesla believes it can meet tomorrow’s production demands no matter who it chooses—and that’s a positive sign for future autonomous development.
Sorry, but there can never be autonomous cars. Most of what this article implies, like Machine Learning, is way too slow, so then has nothing to do with autonomous vehicles. But everyone already knows this. For example, there is no way any computer could ever drive after a snowstorm. It can be hard enough for a human, and we process visual information over 100 million times faster than the fastest computer. But the idea of Tesla going out on its own over a dead end project like autonomous vehicle programming, makes no sense. That would mean abandoning all the existing infrastructure of off the shelf computing, like languages, compilers, assemblers, operaring systems, etc. While I do not like many of the current conventions, autonomous cars can’t be driving this. Makes no sense because most people do not even WANT an autonomous car. Most people LIKE the sense of power that they get from being the driver. So the only thing that remotely makes any sense is that this is actually motivated by something else, like secret military or NASA projects where autonomous activity is really needed and much simpler to implement. The one way autonomous vehicles could be possible is if all roads contained embedded transponders and all vehicles were networked, but that can never happen because it would require all current vehicles to be banned.