Rivian‘s head of chip design has detailed the technical architecture behind the company’s custom autonomy processor, which was unveiled at the EV maker’s inaugural Autonomy and AI Day in December.
Mukund Chavan, Vice President of ASIC Design, joined Rivian in October 2021 after nearly four years at Tesla, where he delivered the Full Self-Driving computer HW3 into volume production across the Model S, Model X, Model 3 and Model Y.
He previously held senior roles at Apple’s Special Projects Group, the name given to the then-scrapped Apple Car Project.
The executive described RAP1 — the in-house developed 5nm chip — in a detailed interview published Thursday on the ‘Creative Strategies, Inc.’ YouTube channel.
5B Pixels Per Sec
The R2’s sensor suite streams 65 megapixels per frame across 11 cameras, supplemented by high-resolution imaging radar and long-range LiDAR.
Two RAP1 processors inside the Autonomy Compute Module 3 together handle 5 billion pixels per second. The perception stack runs on a 100-millisecond loop, vehicle controls operate at 100 hertz and GPS and IMU sensors run in the kilohertz range.
“A typical attentive human has a response time of around 1.5 seconds,” Chavan said. “The autonomy system with its multimodal 360 sensing is able to detect and respond to the environment in about 5 to 10x faster.”
The chip processes camera streams through an on-chip image signal processor handling pixel processing, format conversion, region-of-interest cropping and lens correction.
Radar and LiDAR data arrive as point clouds and feed into Rivian’s Large Driving Model — an end-to-end neural network that generates a world model, tracks vehicle state and handles trajectory planning.
RAP1’s internal subsystems include an always-on processor for secure boot, a dedicated ASIL-D safety island, high-speed sensor interfaces, an in-house neural network engine, 14 ARM Cortex-A720AE general purpose cores and a general purpose GPU. The chip is manufactured on TSMC’s N5A automotive process node.
“We benefit from working on a mature platform so that we can put it into our safety critical systems,” Chavan said.
Why Rivian Built It
Chavan said the decision to design components internally versus licensing them followed a disciplined evaluation of available options.
“We went through a disciplined approach, looked at all the IPs available outside — you have even like neural network accelerators — and we determined that where we can add the most difference is, because we have a full understanding of the hardware software stack and are vertically integrated, we can add the most differentiation by designing our own neural network engine,” he said.
Rivian also designed RivLink, a proprietary low-latency chip-to-chip interconnect, in-house. The interconnect allows multiple RAP1 chips to be connected to scale processing capacity.
“We’ve been around the block enough to know that we don’t want to build a single-purpose chip and we want to make sure it’s extensible to other systems,” Chavan said.
A merchant silicon vendor must serve multiple customers with unknown system requirements, limiting available optimizations, according to Chavan.
As one example, Rivian pulled the memory off the board and into the multi-chip module — eliminating the need for a high-density interconnect board and reducing system cost.
That optimization is only possible because Rivian controls both the chip and the system around it.
Sizing 800 TOPS
Chavan gave a detailed account of how the chip’s compute capacity was determined. Die size comes first, shaped by yield economics on TSMC’s N5A node.
Non-negotiable elements — memory interfaces, high-speed sensor interfaces, power delivery and the sensor pipeline — consume their share of the silicon area.
“We look at our applications and profile them that are running on a previous generation system and we say, let’s pick a target of let’s say 3x the performance and see how we can fit into that,” he said.
“We said okay, let’s give a little bit more headroom and target 4x the performance. So that’s how we arrived at it,” the VP added.
Future-proofing drove the decision to build in headroom, according to the former Tesla and Apple executive.
“You have to be able to put in the compute now so that you can allow for over-the-air updates so that the software team can continue to add features and functionality. A vehicle is going to have a lifetime of several years.”
Chavan said the chip marks a turning point for how Rivian is perceived by the industry and by investors.
“We can claim now that we have silicon that we are truly vertically integrated company that capable of doing more than the space and delivering value to the customer through our products and also advancing Rivian‘s mission and taking RAP1 into production will also put Rivian squarely in the AI space.”
Timeline and Production
From project start to silicon in hand, the timeline was approximately two and a half years. The team spans SOC architecture, design, verification, packaging, signal integrity, power integrity, physical design and design-for-test architecture.
While Rivian‘s Autonomy Day guidance pointed to late 2026 for production integration, the most recent earnings call indicated early 2027.
Integration work remains ongoing.
At December’s event, CEO RJ Scaringe framed RAP1 as a strategic inflection point.
“Our updated hardware platform, which includes our in-house 1600 sparse TOPS inference chip, will enable us to achieve dramatic progress in self-driving to ultimately deliver on our goal of delivering L4,” Scaringe said.
“This represents an inflection point for the ownership experience — ultimately being able to give customers their time back when in the car.”
Software Revenue Accelerates
The chip development comes as Rivian‘s business model shifts increasingly toward software.
Rivian has disclosed Autonomy+ pricing of $2,500 as a one-time purchase or $49.99 per month — below Tesla’s Full Self-Driving offering.
The company expects software and services revenue to grow approximately 60% in 2026, CFO Claire McDonough said at the Morgan Stanley Technology, Media & Telecom Conference on Thursday.
The guidance implies revenue of approximately $2.5 billion this year, up from $1.56 billion in 2025 and $484 million in 2024.
Growth is driven by the Volkswagen Group joint venture and an expanding suite of subscription products.
McDonough confirmed that Rivian will launch paid subscriptions for its Autonomy+ autonomous driving software in April.
“In April, we’ll start our paid subscriptions within the software and services component and see that as being a long-term catalyst for margin expansion opportunity for us as well,” she said.
