At embedded world 2026, TI has launched a new family of real-time microcontrollers (MCUs) with a proprietary on-chip micro-NPU: the AM13E230x.

This is the third MCU in the TI lineup to get the company’s NPU, branded “TinyEngine,” Amichai Ron, senior VP for embedded processing at TI, told EE Times.

“It’s exciting; the microcontroller has been the same for decades,” Ron said. “We now see a revolution at this level of performance that will allow customers to design systems that were not possible five years ago.”

The new family, AM13E230x, is Arm Cortex-M33-based real-time MCUs designed for adaptive motor control and predictive maintenance in appliances, industrial systems, and robotics. The devices support real-time control loops for up to four motors at a time. The TinyEngine NPU can offload AI-based control algorithms or predictive maintenance/fault detection algorithms from the M33, improving latency and power efficiency.

“We can run neural networks on the Arm processor; we don’t need the NPU,” Ron said. “That’s good for certain applications, but where you care about latency, or you want to reduce the overall power consumption, the NPU is a lot more efficient.”

The TinyEngine NPU in the AM13E230x can achieve 2.56 GOPS and supports 8-, 4-, and 2-bit precision, as well as various mixed-precision modes. This is the same NPU already in the F28P55 (launched in November 2024), a C2000 MCU that uses its NPU for motor bearing and solar arc fault detection, offloading from the CPU, which can be dedicated to real-time motor control.

The TinyEngine NPU is also already in some members of the MSPM0 family, a general-purpose MCU with an Arm Cortex-M0+ core. This part consumes less than 2µA in standby mode and can handle applications like wake-word detection, gesture recognition, and predictive maintenance.

“Latency is becoming critical,” Ron said. “You can detect faults with a microcontroller, but it will take you half a second. With the AI acceleration, we can do it in four milliseconds, so you can shut down power to the system as quickly as possible before there’s a safety hazard.”

TI is keeping the architectural details of the TinyEngine under wraps for now. But unlike the C7x, TI’s bigger AI accelerator architecture for its automotive products, which is a mix of DSP and matrix-multiply acceleration, TinyEngine has no DSP component. Its use cases at the sensor edge are very different from the C7x’s in a complex multi-camera ADAS or AV system in a car, Ron said.

TI launched its latest generation of parts with the C7x accelerator at CES—the TDA5 family. This Arm-based SoC family supports up to Level 3 autonomy in AVs and robots, with a dedicated vision subsystem that can handle up to 16 simultaneous camera feeds. As well as the C7x, there are also hardware accelerators for common vision tasks and a “substantial” on-chip memory that reduces the need to go to off-chip DDR to keep latency low.

The TDA5 family scales from 10 to 1200 TOPS (INT8) by using multiple C7x cores on a chip. Big convolutional neural networks and transformer models can run with power efficiency as high as 24 TOPS/W.

This scalability is critical to automotive applications, Ron said.

“A lot of the investment in these applications is on the software side,” he added. “We want to make sure customers don’t have to keep moving the software if they need more or less performance.”

An existing family, TDA4, covers 4-32 TOPS. The version of the C7x in the TDA5 has the same mix of vector DSP and matrix-multiply acceleration, upgraded for power efficiency and ease of use, Ron said.

TI advertises the TDA5 as “chiplet-ready,” which can add another level of scalability. With UCIe on-die, options to collaborate with automotive OEMs are on the table, Ron noted.

“We are going to look at opportunities to collaborate,” he said. “UCIe is a standard interface, so we should be able to collaborate with other industry leaders and to be able to offer more options for our customers. Everything is on the table. We have concrete plans to use [chiplets] ourselves, and with partners is also an option.”

Software stacks Software approaches for TinyEngine and the C7x are understandably different.

“People using the C7x usually have a lot of experience with AI, so we give them access to all the usual tools, and they can figure out what they want to do with the system,” Ron said. “With TinyEngine, most of the customers using it have a more limited knowledge of AI. Of course, we offer all the usual tools, but we also offer Edge AI Studio to collect data, train a model, and compile it to the MCU, so they can do it without a deep understanding of AI.”

TI also offers sensor kits that can be used for data collection. Customers typically begin with AI in fault detection for a particular industrial subsystem, before expanding to multiple subsystems, Ron said.

“Over the next two or three years, understanding [of AI among embedded developers] will accelerate significantly, especially as we introduce more devices and more tools, and our competition does the same,” he said.

As well as providing CCStudio Edge AI as a royalty-free development environment, TI works with third-party development environments for AI, enabling customer choice, Ron said.

“The whole [MCU] industry is working together,” he said. “The average engineer doesn’t know how to use AI yet. Our job [as an industry] is to offer these capabilities and train more people to use edge AI capabilities that didn’t exist just a few years ago. The whole industry is driving in this direction, and it’s a very exciting time.”

Ultimately, TI’s goal is to offer parts up and down its portfolio in versions with and without AI acceleration, compatible with TI tools and any third-party tools developers like, Ron said.

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved