Jun-sik Park, Country Head (left), and Ted Lim, Senior Manager

As the SDV narrative narrows toward central HPCs, it becomes easy to forget the time in which a vehicle actually runs. ST’s Stellar P3E is not about showing off AI performance; it is about inference that must finish within a defined deadline - inside physical loops of power, heat, torque, and braking. In the same way, integration (x-in-1) is not merely an event of bundling parts. In an era where platform variants and OTA repeat, it is a decision to reallocate time across validation, safety, and security. The question is not “How fast is AI?” but “Where should AI be placed to make the system stronger?”

By Sang Min Han _ han@autoelectronics.co.kr
한글로보기






In late February, ST introduced its automotive MCU “Stellar P3E,” equipped with AI acceleration. Jun-sik Park, Head of ST Korea, is not someone who often takes the stage at press briefings. That is why his first line was not about the “product” Stellar P3E, but about “direction.”
Park framed ST’s automotive strategy around smart mobility, power & energy, and cloud/connected/autonomous - then pinned the focus to the heart of electrification. Behind that focus are three overlapping pressures: safety, EVs, and internal combustion efficiency. These are not challenges from different eras; they stack inside the same vehicle at the same time. As pressures stack, in-vehicle electronics become more complex - and as complexity rises, the decisive factor moves from “faster computation” to computation that always finishes within a fixed deadline: determinism.
AI follows the same logic. AI is no longer only a story of central computers. The moment AI moves down into the time domain of “physical loops” - where motors spin, battery thermal states fluctuate, and high-voltage switching repeats - the SDV center of gravity shifts back toward the MCU. If the AI of central HPCs speaks the language of perception and inference, then the AI of the propulsion domain must be inference that completes inside the control loop to have meaning. Stellar P3E is one of the clearest chips to show this message. It is labeled an “AI-accelerated MCU,” but the real question ST posed was: Where does AI belong for it to matter?
Ted Lim, Automotive MCU Marketing, presented a timeline for Stellar P3E: a “mass-production target by the end of this year,” and vehicles in which it is deployed “around 2028 - 2029.” This was not merely a product announcement - it was a preview of where the center of electrified integrated control will move in the coming years.
The core of that direction is that integration and time move together. x-in-1 integration bundles functions for performance, efficiency, and cost - but from that point on, the essence of the fight shifts to determinism and reliability. And when a world of platform variants and ever-changing requirements via OTA overlaps, the system must be flexible yet robust. It must accept expansion and change without shaking the fundamentals of safety, validation, and security. That is exactly why Park began with “direction,” not “product.”






The Place of P3E in an SDV Architecture:
“Not an HPC - A Physical Computer That Integrates Propulsion”

When people talk about SDV today, attention typically gravitates toward central HPCs for ADAS and IVI. But the SDV architecture Lim presented carried a different implication. At the top sits an HPC for ADAS and infotainment. Beneath it is what one could call a physical computer, responsible for propulsion and driving. Below that sit zone/domain controllers and x-in-1 integrated controllers, and at the edge, smaller MCUs driving sensors and actuators. A “physical computer” is the computing layer that closes physical quantities - power, heat, torque, braking - at fixed cycles.
In ST’s explanation, P3E sits in the electrified integrated controller (x-in-1) domain - more specifically, the MCU for a “domain controller” in the powertrain/energy domain that integrates and manages the inverter, converter, OBC, DC-DC, and BMS. In other words, the chip at the heart of an EV.
“Our market is the MCU market. SoC products from companies like NVIDIA or Qualcomm focus on higher performance - video or ADAS. ST’s products target integrated controllers. SoC or processor types tend to be weaker in real-time behavior, which is why you end up using embedded MCUs,” Lim said.
Real-time behavior - especially in propulsion and energy domains - must be deterministic. Deterministic behavior matters: if a task must be done within 1 ms, it must always be done within 1 ms. More precisely, not on average, but even in the worst case. If latency fluctuates, “fast computation” does not help control.
This is close to the reason P3E exists. Even as SDV becomes “software-centric,” the vehicle must still run: current must flow, motors must spin, and heat must be managed. Inside this physical loop, what matters is not only “fast calculation,” but calculation that reliably completes within a defined time - and now, ST is saying AI should be inserted into that same deadline.



Why Integration Is Inevitable
Harness, Weight, and Co트리플 슬롯 - The Pedal of Electrification

The way ST explained why integration is inevitable in electrification was also notable. Raising EV efficiency is not simply a matter of increasing battery capacity. You need to reduce system weight, minimize losses, shrink packaging, and cut the number of parts. The starting point is the wiring harness. Reduce the harness and you reduce weight; reduce weight and you can travel farther with the same battery - or achieve the same performance with a smaller battery. In the end, the entire cost structure shifts. Lim described this as “optimization of the vehicle wiring network,” including both power and communications - and the x-in-1 integrated controller is a way to maximize that optimization.
At the OEM level, integration is not merely combining parts. It is a choice that changes how the platform is designed, produced, validated, and serviced. Fewer harnesses and connectors reduce assembly steps and lower quality risks such as fastening errors, mis-mating, and contact failures. Reduced volume creates packaging headroom and can shorten system design time across cooling, shielding, and service accessibility. But as modules consolidate, functions and responsibilities concentrate into a single block - so the burden of validation, safety, and updates also concentrates into a single point. That is why integration is not a “hardware event” of reducing ECU count; it is a decision about how to reduce the time cost of development, validation, and updates in an era of repeated variants and OTA.
ST emphasized the effect of its x-in-1 integration solution in strong numbers. Compared to conventional approaches, weight becomes 43% lighter, volume 27% smaller, and high-voltage efficiency improves by 1%. It also cl트리플 슬롯med that high-voltage connection-area costs can be reduced by 60%, and that integration lowers development complexity, reducing required software/tool demands by 75%.
The key is not the absolute values but the direction: integration disrupts cost, packaging, and the development system simultaneously. ST’s logic connects as follows: integration comes first; as integration grows, control and validation become harder; and AI enters to reduce that difficulty - through energy optimization, condition diagnosis, and calibration automation.
As integration grows, the number of individual MCU units may decline. But from a portfolio perspective, the larger MCU (P/G) responsible for integration sees its scope, safety level, and network role expand - while smaller edge MCUs remain necessary. ECU counts may fall, but the responsibility carried by a single MCU grows. ST is effectively covering both a “shrinking area” and an “expanding area” at the same time.







트리플 슬롯 Buys Time Inside the Control Loop:
The Neural-ART Accelerator

Integration comes first; as integration grows, control becomes more difficult. And AI enters to finish that harder control - within time. The need for an AI accelerator is not only about safety assistance. In an electrified integrated controller, AI’s primary purpose is efficiency: reducing inverter switching losses, smoothing motor torque, estimating battery SOC and thermal states more precisely to optimize power and cooling. All of these computations must finish inside the control loop. As integration grows, the room for efficiency gains expands - but time constraints become tighter. Dedicated accelerators like Neural-ART bring inference into the control cycle.
So what “feel” does an AI accelerator create in a vehicle? ST answered with an example. Using window anti-pinch as a case, ST compared running the same AI model on the CPU versus on P3E’s AI accelerator - showing 69× faster response in object-pinch detection inference and 16× faster response in determining full closure, with lower power consumption.
“Windows are just an example. For propulsion - like a traction inverter - you must use AI inside the control loop of a motor rotating at extremely high speed. It’s hard to run that on the host core, so you need an accelerator to shorten inference time to apply it to the system,” the company explained.
What AI does in the powertrain was also described with relatively concrete examples: smoothing the switching region of power semiconductors, driving motors more efficiently to reduce battery consumption, estimating battery SOC/SOH more precisely to optimize charging/discharging and thermal management. Control moves from rule-based approaches - like “turn the HVAC on when temperature exceeds X degrees” - to data-driven optimization.
AI also connects to cost reduction. If a function previously required sensors, and AI can infer the same condition from motor feedback alone, you can reduce sensors. AI is computation, but it also changes BOM structure. This is not AI that is lumped into central computers, but AI that runs “small, fast, and deterministically” close to physical loops - and that AI influences the cost and structure of integrated controllers. From the OEM perspective, the value of AI is not “performance display,” but whether you can increase computation inside an integrated module without losing time - while also reshaping system cost and structure.
Here, real-time behavior returns to the fundamentals of the MCU: how latency is controlled, how pipelines are designed, and why an automotive core was chosen. The answer is simple - because it was designed that way. That is why the MCU rem트리플 슬롯ns essential.







xMemory Targets Development Speed as Much as Performance
Reducing the Trap of Variants

Another axis of P3E is memory. 트리플 슬롯 calls this xMemory and positions PCM (phase-change memory) as the core foundation. While flash has functioned as the de facto 트리플 슬롯andard in the MCU space, 트리플 슬롯 adopted PCM at the 28-nm generation to expand memory density and scaling headroom. As SDV progresses, platform variants repeat and requirements and validation burdens grow; in that context, PCM touches not only performance but development speed and the co트리플 슬롯 of iteration.
“When an OEM develops one product, it then branches into a first, second, third model. As it branches, segments change, requirements shift slightly, development happens again, memory inevitably grows, and in the end you have to use different silicon. Each time, additional resources must be allocated.” Lim said.
xMemory argues that with enough memory headroom, you can cover increased requirements in derived models within the same family. If you don’t have to jump to different silicon, migration, validation, and redevelopment costs decrease.
For a representative P3E configuration (P3H3), the memory figures presented were up to 19.5 MB of xMemory and up to 1.8 MB of RAM. This is not merely a spec sheet; it connects to the logic of reducing the development-resource trap that grows as variants repeat. Ultimately, ST’s competitiveness is not whether it offers “SoC-class compute,” but whether it can make development and validation of electrified integrated controllers less painful. For OEMs, it works in the direction of reducing the repetitive development and validation time that accumulates each time a platform expands.



Making an Integrated Controller “Safe”
Analog/Timer/IO Efficiency and High-Speed Networking

As integration grows, systems become riskier - and the electrification domain is especially so. Inverters, converters, OBC, DC-DC, BMS - these are the lifelines of an EV. That is why ST, after talking about integration, quickly moved to “how to integrate safely.”
Lim emphasized precise timers, rich analog resources, and IO efficiency. The precision of motor control and thermal management ultimately depends on analog interfaces and timing. P3E significantly expands analog resources - ADCs and comparators - t트리플 슬롯lored to electrification dom트리플 슬롯ns.
ST presented a configuration built around 106 ADC channels, including 12× SAR ADC, 10× Sigma-Delta ADC, and 4× comparators. The point was to show, in hardware resource allocation, the ability to measure precisely and control precisely in an electrified integrated controller.
Networking was also pulled into the SDV direction with higher speed. P3E points toward MCU-level support for Ethernet-based real-time networking such as TSN (Time-Sensitive Networking) through 10/100/1000BASE-T1 Ethernet, 10BASE-T1S, CAN XL, and CAN FD. The key is not bandwidth itself but predictable latency.
In this flow, safety and security are not optional - they are preconditions for integration. ST emphasized ASIL-D aspirations and hardware-based security (HSM), describing “360-degree security.” As integrated controllers grow, a single fault carries a larger impact. Integration reduces parts, but it also reallocates the OEM’s time exposure to validation, accountability, and recall risk.



Timeline: Mass-Production Target by Year-End, Vehicles in 2028 - 2029
The Deciding Factor Is the OEM’s “Will to Integrate”

“We have already started sampling, and development is underway with major OEMs and Tier-1s after providing samples. The production target is the end of this year, and you will likely see vehicles that deploy it around 2028 - 2029,” Lim said.
This schedule is less about “when it arrives” and more about when market collision begins in earnest. The contest will not be only about whether an AI accelerator exists, but about how far OEMs push x-in-1 integration. Some may stop at 3-in-1; others may push beyond 5-in-1 and 7-in-1 to more aggressive declarations like 14-in-1. Depending on that choice, the size and role of the integrated controller - and the way AI is inserted - will change.

As SDV expands, integration does not stop. Inverter, converter, OBC, DC-DC, BMS are bundled into a single electrified integrated controller. Data paths and power paths, and software accumulated through OTA, all converge. Integration simplifies structure while placing more computation and more responsibility onto one chip. So the decisive factor becomes not “how much you compute,” but whether you can finish within time.
The essence of what ST calls an “AI-accelerated MCU” is increasing computation inside an integrated structure without shaking the time of the physical loop - this is determinism and reliability. That is also why Park began with “direction,” not “product.” ST is pulling that role back to the MCU (edge) and is targeting a doubling of automotive MCU revenue by 2030 compared to 2024.

AEM(오토모티브일렉트로닉스매거진)