Munich Exhibition Frontline Observation: What's Happening in the MCU Industry?

In early July 2026 in Shanghai, at the Munich Shanghai Electronics Show, the booths of embedded manufacturers were busier than ever. Engineers visiting the exhibition were no longer only asking about prices and delivery times; more people stopped by demos of dexterous hands for humanoid robots and edge AI development boards. So, where is the trend in the embedded industry this year?

How is AI MCU Adoption Going?

The concept of "AI MCU" has been around for a long time, but how is the actual adoption?

Texas Instruments (TI) showcased an edge AI digital recognition solution integrating the TinyEngine NPU on site. The TinyEngine NPU is a dedicated hardware accelerator launched by TI, providing 2.56 GOPS of computing performance, specifically designed for deep learning and inference operations. However, industry insiders say that the computing power of AI MCUs generally falls in the range of a few tenths of a TOPS, only capable of running small neural networks with tens to hundreds of thousands of parameters, mainly used for "classification," such as millimeter-wave radar for gesture classification or anomaly detection in motor vibration signals. TI staff admitted that these types of embedded devices are not for centralized processing but for "small tasks" in specific scenarios. They also launched a foolproof IDE tool, claiming to "guide you step by step" through data collection, labeling, training, and deployment, even directly outputting header files and compiling network weights into the project. More notably, TI has integrated AI into its own CC Studio integrated development environment, directly linking to large models to achieve a one-stop service of "requirement input → automatic coding → automatic programming → automatic debugging." As one engineer at the scene put it, "You basically don't need programming skills anymore; just being able to understand things is enough." In fact, the functions on an MCU are not particularly complex, and AI-assisted programming is most advantageous in "simple scenarios." After all, if the task is too complex, the MCU's memory and computing power simply can't handle it, limiting its potential.

STMicroelectronics (ST) brought the STM32N6 edge AI MCU chip, with 0.6 TOPS of computing power, featuring a self-developed NPU, running standard CV models capable of gesture recognition and keypoint detection. Chen Deyong, from ST China's Microcontroller Product Division, emphasized in a concurrent report that edge AI deployment must meet "two smalls and two lows": small Flash, small RAM, low power consumption, and low latency. He also revealed that ST started developing NPU test chips as early as 2015, and now has over 5,000 MCU/MPU products based on the ARM platform. It is also the only manufacturer globally to provide an online platform for MCU benchmarking, with 73% of ML Perf Tiny developers choosing STM32. However, AI MCUs also face the issue of precision loss during model conversion from "floating point to integer." ST's recommendation is: during the training phase, first compress the model using floating point "to the extreme," then convert to Integer 8 for deployment, balancing Flash usage and inference time.

ARM promoted "Zephyr RTOS" in a technical report, aiming to open a migration path from Cortex-M to Cortex-A. ARM believes that edge AI requires greater computing power expansion, but developers do not want to jump from a real-time operating system to a bloated Linux. Zephyr retains the lightweight and predictability of RTOS while leveraging the MMU and multi-core SMP of Cortex-A, enabling smooth upgrades for embedded code. ARM also mentioned on site that its Ethos-U series NPU (U55/U65/U85) can now be used with Cortex-M and Cortex-A, and TensorFlow Lite and PyTorch runtimes are already ready.

However, the adoption of AI MCUs is still stuck on "data collection" and "regulations." TI staff mentioned that the 60 GHz millimeter-wave radar used for AI MCU data collection has not yet had its standards clearly opened in China. "It hasn't been said that it can't be used, nor that it can be used," but competitors could use this to file complaints, making industrial customers generally cautious. Meanwhile, camera-based data collection solutions raise privacy concerns. LiDAR is expensive, costing thousands of yuan per unit, while an AI MCU only costs a few dozen yuan. In terms of application directions, TI also mentioned potential future scenarios such as elderly care and health care, but currently, such demand has not truly been unleashed. A common judgment among multiple manufacturers is that edge-side AI MCUs are currently "lukewarm," with a real explosion still dependent on further maturity of application scenarios and regulations.

Are Humanoid Robot MCUs a New Blue Ocean?

If AI MCUs are still waiting for the wind, humanoid robot MCUs are already standing at the threshold from pre-research to small-scale production.

One of the most attention-grabbing demos at the show was a six-degree-of-freedom dexterous hand exhibited at ST's booth. Each finger base is driven by a coreless brushless motor, with finger bending driving the upper joints to move in coordination, capable of briefly gripping a dozen kilograms. ST staff explained the chip solution inside: the fingers are controlled by an M4-core MCU, the overall hand is managed by an M33-core MCU, and the next version will add an N6 chip for edge perception. In the entire solution, the chip cost does not exceed 100 yuan, but the true cost driver is the motor—one coreless brushless motor costs 700-800 yuan, with six motors in one hand, and the entire robot has even more motors, far exceeding the MCU cost. It is understood that some of ST's MCU products for humanoid robots are already being manufactured locally in China through partnerships with domestic foundries, and even exported overseas. Staff said that the humanoid robot supply chain in China is already very mature, making it impossible for foreign costs to come down; most components for Tesla's robots are also sourced from China.

GigaDevice's booth showcased humanoid robot-related solutions, with its comprehensive layout and deep implementation standing out at this exhibition. The on-site display featured four cutting-edge control solutions for high-degree-of-freedom robots: first, a six-axis robotic arm solution based on the GD32H75E; second, a "Gang Beng L1" quadruped robot developed in collaboration with Zhishen Technology, deeply integrating multiple GD32 MCUs internally; third, a robot joint drive solution also based on the GD32H75E; and fourth, a six-dimensional force detection solution based on the GD30AD3642. Staff revealed that different parts of the robot have very different requirements for MCUs: lower limb joints use planetary reducers, emphasizing high power, high-temperature resistance, and long life, with less emphasis on precision; upper limbs and dexterous hands require high precision, and in the future, even surgical- or massage-level micro-operations will be supported. A single bipedal humanoid robot may require hundreds of MCUs, with higher degrees of freedom requiring more.

Geehy Semiconductor showcased an even more extreme cost-performance route. Their APM32M3514 chip (M0+ core, 72 MHz) for robot motor control costs only about 3 yuan per unit, while the higher-performance APM32F425 (M4F core, 240 MHz) is used for shoulders and legs, paired with a dedicated MCU (M52 core, 128 MHz) G32R430 encoder to obtain motor angles. Geehy admitted that most humanoid robot projects are currently in the pre-research phase, with mass production still needing time, but chip manufacturers must position themselves early.

Infineon demonstrated another technical path: a 1kV GaN joint motor drive solution based on the PSOC Control C3 MCU. This solution integrates main control, power, and sensing, utilizing the high-frequency characteristics of GaN devices to push the switching frequency to 100 kHz, far higher than traditional silicon-based solutions, thereby reducing switching losses and heat generation at the same power level. This means robot joints can achieve higher power density in a smaller volume, which is especially important for high-power joints like those in the lower limbs. Infineon chose to synergize wide bandgap semiconductors (GaN) with dedicated MCUs for higher energy efficiency.

The control of humanoid robots depends not only on the computing power of individual MCUs but also on the communication efficiency of the MCU network. NXP's showcased i.MX RT1180 solution connects the main MCU to five motor drive MCUs via the I3C bus, achieving 12.5 Mbps synchronous communication with just two signal lines, eliminating the need for an external crystal oscillator to save BOM and PCB space on the slave side. This solution connects from the dexterous hand's local I3C bus to the entire body's EtherCAT, meaning that when hundreds of MCUs are deployed throughout the robot, local control and overall main control can collaborate through a high-bandwidth, low-wire-count standardized bus, rather than relying on point-to-point asynchronous communication stacking.

Beyond motion control and communication for dexterous hands, upgrades in the perception layer are equally critical. ADI's multimodal tactile sensing solution integrates a 32×32 high-density tactile array with edge AI inference. The tactile network can synchronously sense pressure distribution, micro-vibrations, contact status, and temperature at kHz-level frame rates. The robot dexterous hand no longer relies solely on visual open-loop control but can achieve precise gripping of fragile objects, anti-slip control, and fine-tolerance assembly through tactile closed-loop feedback.

During interviews, multiple manufacturers reached a consensus: current robot shipments are still very small. Even if Unitree sells tens of thousands of units a year, for chip manufacturers, that's only "a few million units," while their monthly shipments are often in the billions. So, making MCUs for robots is currently "not profitable"—it's purely betting on the future. The true cost of humanoid robots is not the MCUs but the reducers—since specifications are not yet standardized, each company develops its own, resulting in no economies of scale and persistently high prices. ST even stated that the robot business is currently "aiming for the future, hard to ramp up volume now," with MCUs following a "high volume, thin margin" model.

Can RISC-V Replace the ARM Architecture in the Embedded Space?

Beyond the ARM ecosystem, RISC-V is another increasingly clear MCU route at this year's exhibition.

WCH (Nanjing Qinheng Microelectronics)'s "Qingke" MCU is a typical representative, focusing on interface chips. They fully transitioned to RISC-V as early as 2018, developing their own cores, and now rarely ship ARM products. The on-site engineers gave straightforward reasons: first, security—free from geopolitical influences; second, cost—directly saving customers money; third, performance is not bad—the CH32V203 is comparable to the STM32F103, with advantages in CoreMark scores and power consumption.

Silergy also took the RISC-V path. Its SA32D series MCUs are based on high-performance, high-reliability RISC-V cores, with ultra-high computing performance and rich peripherals, mainly used for high-safety main control edge computing and control scenarios such as zone control (ZCU), battery management system (BMU) for power batteries, main drive motor control, chassis applications, and some ADAS applications. The products comply with the AEC-Q100 standard and functional safety ISO 26262 ASIL-D. Silergy stated that RISC-V IP forms are now abundant, from M0-level to 64-bit server-level, and the ecosystem is spreading from consumer → industrial → automotive. Regarding the "AI MCU" concept, Silergy's attitude is relatively calm: everyone is watching, but the NPU computing power on MCUs is only 0.3-0.5 TOPS. "What can such weak computing power do? The key is whether the scenario can bring value."

MRAM to Become a New Solution for Future Embedded Storage

The evolution of embedded microsystems has always revolved around performance, cost, and reliability. The continuous upgrading of scenarios such as edge AI inference, real-time robot joint control, and high-frequency industrial sampling is not only reshaping the core of microsystems—the MCU itself—but also forcing the entire storage system to iterate simultaneously. Traditional Flash, limited by erase/write cycles, write speed, and block operation modes, is already unable to meet the multiple requirements of high speed, high endurance, and high reliability at the same time. As a representative of new non-volatile memory, MRAM is becoming a new technical path to support the performance upgrade of embedded microsystems.

At this exhibition, domestic manufacturer Zhejiang Chi Tuo Technology Co., Ltd. brought its MRAM-related solutions. It has laid out a 12-inch MRAM production line in Qingshan Lake, Lin'an, Hangzhou, and is currently the only manufacturer in China to achieve mass production of STT-MRAM. Its current main products have read/write speeds of tens of nanoseconds, support bit-by-bit read/write without erasure, are non-volatile with power loss, and have endurance up to trillions of cycles, far exceeding Flash's microsecond-level write speed and hundred-thousand-cycle endurance. The next-generation technology, SOT-MRAM, is in the research and development stage, with read/write speeds potentially reaching nanosecond levels, and related results have been published in the top journal IEDM. In terms of capacity, current standalone chips cover capacities from Kb to 64 Mb, with plans to evolve to 128 Mb to Gb levels. Another important business is embedded MRAM (eMRAM)—integrating MRAM directly into MCUs/SoCs to replace traditional embedded Flash or cache. Chi Tuo has already cooperated with leading domestic MCU companies.

It is understood that in terms of application scenarios, compared to the SRAM+battery power-loss protection solution, MRAM does not require a battery, completely solving maintenance issues in explosion-proof and remote scenarios. More notably, as traditional memory like DRAM prices continue to skyrocket, MRAM can also reach the SCM (storage-class memory) market. In scenarios extremely sensitive to power consumption, such as wearable devices, MRAM's non-volatile nature eliminates the need for continuous power supply like SRAM or DRAM, naturally offering low-power advantages. Compared to another new memory type, PCM (phase-change memory), MRAM excels in wide temperature range, making it more deterministic in industrial-grade and automotive-grade applications. In the industrial field, its products have entered areas such as industrial PLC/DCS and commercial energy storage, with customers including leading domestic industrial control companies and energy storage companies. Its representative products have passed AECQ100-Grade1 automotive certification and are being promoted to automakers and Tier 1/2 suppliers.

Currently, Chi Tuo's embedded MRAM products are still catching up with international leaders like TSMC and Samsung, but its standalone MRAM products can compete on par with foreign companies. In addition to its own products, it also provides tape-out services to universities, research institutes, and startup semiconductor companies, making it a rare domestic platform for new memory and micro-nano manufacturing.

Chi Tuo believes that with the evolution of in-memory computing technology, MRAM may play a greater role in AI chips in the future. Currently, it starts by entering scenarios such as industry, energy storage, and general transportation, gradually expanding to the AI field.

Looking back, the competitive focus in embedded systems is shifting toward parameters—who can better accurately match scenarios, who can more independently control the architecture, and who can more effectively restructure the storage foundation. The 0.x TOPS computing power of AI MCUs is not impressive, but it is sufficient to drive intelligence in niche scenarios. The hundreds of MCUs used in humanoid robots are not exaggerated, but they are enough to reshape the chip demand curve for motion control. The IP licensing of RISC-V is not complicated, but it is enough to break the single dependence on the ARM ecosystem. The capacity of MRAM is not yet high, but it is enough to establish new standards for non-volatile storage in industrial and automotive fields. Individually, these changes are not dramatic, but taken together, they are redefining the boundaries of embedded systems.

Source: Semiconductor Industry Vertical and Horizontal

Risk Warning and Disclaimer

        Market risk exists; investments should be made cautiously. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situations, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investment based on this is at your own risk.