HWNEWS
Cars are consolidating onto a few powerful central processors.
Cars used to be wired with dozens of small, dedicated computers: one for the engine, one for the dashboard, one for the brakes, another for the radio. Over the past decade, automakers have steadily replaced that sprawl with a few powerful central processors that now run the dashboard, driver-assist, the network gateway, and core vehicle control on a single platform. This consolidation is what makes the "software-defined vehicle" possible, and, according to an industry analysis of the shift, it is also what is running out of room.
The new central computers are being asked to handle AI workloads that barely existed a few product generations ago. Neural processing has moved from fractions of a TOPS to tens, then hundreds of TOPS inside the same hardware class. CPU performance has jumped from tens of thousands of DMIPS into the hundreds of thousands, and GPU throughput has climbed by orders of magnitude, a curve that, as the same analysis lays out, no longer fits the trajectory of conventional monolithic scaling. At the same time, per-generation power savings from smaller process nodes have shrunk, so efficiency gains no longer cancel out the demand curve. High-end monolithic system-on-chips are now hitting die-size limits, yield problems, and cost curves that no longer pay back, and LPDDR memory tops out around 512 bits of interface width and a few hundred gigabytes per second of bandwidth, not enough, the analysis argues, to feed a centralized AI workload indefinitely.
The industry's structural answer is to stop trying to build one giant chip. Designers are now breaking the central computer into smaller dies, called chiplets. Those dies are integrated inside one package, the way a Lego model is built from separate bricks that snap together. Different dies can be made on different process nodes and mixed as workloads demand, which is supposed to keep cost, performance, and flexibility in balance. Two design choices dominate the roadmap: organic versus advanced 2.5D and 3D packaging, and how to knit the dies together with a unified memory space and cache-coherent interconnect so the software can treat the package as one machine.
That last point is where the bet gets fragile. For cars, chiplets are not just a packaging trick. They have to deliver three things at the same time: cost efficiency, deterministic real-time performance, and long-term software stability, or the architecture does not pencil out. The analysis treats software stability as the deciding factor for whether chiplets succeed in the car, and flags standardization efforts such as UCIe as still working through gaps that matter specifically for automotive adopters. That leaves each platform team to solve a piece of the integration puzzle on its own.
The trade-off is sharper than it looks. A modern central processor must feel like one computer to the software running on it, even when it is physically built from several. If the glue between chiplets is loose, latency and software complexity both rise, and the cost savings from mixing dies evaporate. If the glue is too tight, the design becomes a custom package no one else can reuse, and the whole point of a modular architecture is lost. Organic packaging is cheaper and works for simpler splits; advanced 2.5D and 3D packaging offers tighter integration at higher cost and supply-chain risk. Most platform teams are expected to bet on a "platform-first" scaling strategy, where the package layout and software interface are designed once and reused across several vehicle programs, with new chiplets swapped in as workloads change.
For drivers, the visible result is cars that get better at driving, sensing, and updating over their lifetime, because the same hardware platform can host new AI features without redesigning the wiring harness. For automakers and their suppliers, the visible result is which companies end up controlling the software stack that ties the chiplets together, and whether the per-vehicle cost of that integration can be paid back across multiple model years on the same platform.
The next tell is the first OEM that ships a volume program where a new safety-critical feature lands on existing hardware through a chiplet swap, rather than a new board. That is when the modular bet will have stopped being a roadmap slide.