Physical AI Pushes Hardware to Its Limits with Memory, Power, and Architecture Challenges

icon币界网
Share
AI summary iconSummary
Physical AI systems are transforming industries, but liquidity and crypto markets are encountering new challenges as memory, power, and architectural constraints strain hardware design. Bitjie notes that chipmakers must redesign systems to handle edge AI, balancing compute, I/O, and security. Strategy& forecasts the physical AI market could reach €43 billion by 2030. CFT regulations increase pressure, as custom hardware and software solutions become essential for reliability and safety in uncertain environments.
CoinDesk reports:

Autonomous machines are expected to transform multiple industries, but memory bottlenecks, power consumption limits, specialized architectures, and security requirements are forcing chip designers to rethink the entire system.

Key points:

  • More computing power alone cannot solve the bottlenecks of physical AI. Processing capacity must be matched with memory bandwidth, efficient interfaces, adequate power supply, and cooling capabilities.
  • Physical AI requires a system architecture tailored to its intended function and environment.
  • True implementation depends on rigorous validation: covering edge cases, understanding system behavior, and verifying security.

Physical AI is moving toward widespread adoption, but engineering teams must simultaneously and continuously address multiple technical challenges to build these sensor-driven systems.

Analysts expect the physical AI market to grow significantly. Autonomous systems are poised to fundamentally transform industries such as transportation, manufacturing, healthcare, and energy. PwC’s Strategy& forecasts that the global physical AI market could reach approximately €430 billion by 2030, reflecting the commercial potential of autonomous systems across multiple sectors.

Licinio Sousa, Senior Director of Automotive and Physical AI Interface IP at Synopsys, said: “Physical AI systems are expanding from automotive to robotics and drones, and development bottlenecks for these applications span the entire system. AI is moving from data centers to the edge; for machines such as robots, drones, and vehicles, the driving forces are autonomy, privacy, personalization, faster response times, and reduced reliance on the cloud.”

The list of technical challenges to address is long and deeply interconnected. Memory capacity and bandwidth must be increased, latency and power consumption must be reduced, product lifespan must be extended, and security must be validated. Even more challenging, solving any one of these issues may introduce new problems elsewhere.

Dedicated architecture

Addressing these system-level bottlenecks requires highly specialized, use-case-optimized hardware and software.

Michal Siwinski, Executive Vice President, Chief Product and Marketing Officer, and General Manager of Security Solutions at Arteris, said: “We need to start talking about more optimized, customized computing designed for specific workloads—that’s where physical AI gets really interesting. Those workloads must be highly optimized for hardware. We’re essentially returning to the traditional meaning of ASICs—ASICs were meant to be highly application-specific and extremely optimized, not general-purpose. Trends like this are already emerging. It will be fascinating to see how more flexible systems, tailored for very specific applications, evolve between software, hardware, and workloads—and continue evolving, since models are changing much faster.”

Vasanthan Karunanithy, System Architect for Micron Technology’s automotive business, stated at the recent JEDEC Automotive Forum: “Robots are not a new AI stack. From ADAS systems, robots inherit real-time perception, path planning, and motion control. From smart cabinets, they inherit natural language understanding, contextual reasoning, and human-machine interaction. In embodied intelligence, the greater challenge is running both types of workloads simultaneously and then exchanging information at the end. This is the commonality between these two domains. But there are also differences, reflected in scale of operation. Autonomous vehicles operate in relatively structured road environments and draw power from a larger source—the vehicle battery. However, robots operate in unstructured three-dimensional environments. [Robots] must also maintain multiple degrees of freedom, manipulate objects, maintain balance, and interpret commands within instructions. Compared to cars, they must accomplish all of this with significantly more limited power.”

These differences vary significantly across applications. Autonomous vehicles operate within existing traffic infrastructure but must handle unpredictable road conditions. In contrast, mobile and humanoid robots may need to navigate less structured environments, coordinate manipulation with balance, and operate under more constrained power budgets. Meanwhile, industrial robots typically operate in controlled environments and are connected to fixed power sources. These variations make it difficult to optimize a single architecture for all forms of physical AI.

Designers must build the system architecture on these requirements from the very beginning.

Sousa said: “As workloads and models continue to evolve, we’ve seen the emergence of specialized silicon designed for the specific needs of each class of machine. Our customers are consistently addressing the demand for advanced VLA (vision-language-action) models. At the hardware level, this means more intensive computation and AI engines. Meanwhile, the breadth of use cases makes a single monolithic architecture increasingly difficult to apply, so the industry is shifting toward chiplet designs, where compute, AI, memory, and I/O can be scaled according to specific use cases via the UCIe interface.”

Chiplets offer a potential path for scalable physical AI platforms, especially when heterogeneous computing, memory, and I/O requirements justify advanced packaging. However, they may also increase cost and complexity due to custom packaging, thermal management, and die-to-die communication. For smaller, highly integrated systems constrained by power, size, and cost, monolithic SoCs may be more suitable.

These constraints apply equally to hardware and software. “Sensors will become more use-case focused—you need to ensure you have strong, customized use cases, and you also need a software layer on top,” says Sathish Balasubramanian, Head of Product for EDA AI and Solido at Siemens EDA. “I always joke that everything in the real world is analog, and we’re trying to abstract it into digital, so one of the most important things is redundancy. Sensors are already quite capable—we can sense vision, heat, pressure, and anything else. We’ve also done well with radar and other technologies. More importantly, you need to ensure sensors can operate reliably in their environment, capturing it in a way that doesn’t truly degrade—or by building in redundancy.”

However, the effect of redundancy varies depending on whether components can fail independently and whether the system can detect and respond to failures. Simply adding more sensors does not guarantee reliable operation, especially when environmental conditions may simultaneously affect multiple sensors.

Customizing the computing architecture based on workloads or use cases also involves trade-offs. “How close intelligence needs to be to where perception actually occurs is a challenge with competing demands,” says Yasmine King, Vice President and Head of Automotive at Analog Devices. “Distributing intelligence throughout the system makes updates more complex. Centralization allows for OTA updates through a single point, making patches and fixes easier—but it can introduce latency when rapid response to data is required. In a vehicle, if a pedestrian suddenly steps into the road, you need to respond extremely quickly. So, where you place the intelligence depends on what kind of sensor data you want to collect and how fast you need to react.”

Yasmine King spoke at Reuters Automotive 2026 on October 27, 2026.

Intelligent placement is an architectural decision that involves more than just computational performance. It requires balancing response time, communication bandwidth, power consumption, upgradability, and coordination among distributed components. Centralized computing may simplify management, while local processing can reduce communication latency and preserve critical functions during connectivity outages.

The memory wall of physical AI

Memory capacity, bandwidth, latency, and energy consumption are key bottlenecks for physical AI; certain memory technologies may also face supply constraints. The real challenge is not simply adding more memory devices, but designing memory subsystems capable of supplying data to increasingly demanding AI workloads within strict power and thermal limits.

Sousa said: “Billions of related model parameters also need to be stored and transferred between storage, memory, and processing units. Today, this is known as the ‘memory wall.’ As memory usage continues to grow, DRAM requires ultra-high-speed interfaces, such as LPDDR6, to no longer become a processing bottleneck.”

Performance also depends on access patterns, latency, data reuse, cache capacity, memory controller efficiency, and the energy required to move data. If a workload is limited by other parts of the memory hierarchy, simply increasing raw bandwidth may yield limited benefits.

System design tailored to workloads can help alleviate some memory burdens. “The right-sized memory, the right-sized computation and redundancy, and everything that must fit within thermal and power budgets will be critical,” said Balasubramanian. “Above all, size matters. You wouldn’t put a 100-gallon fuel tank in a Prius—you need a 10-gallon tank. I use this analogy to illustrate memory: you can manage it in a way that doesn’t require massive DRAM, but instead features a robust cache and integrated SRAM. It can always keep critical demands in DRAM, but when needed, you can move other data to flash or SRAM. From a memory path perspective, moving from DRAM to SRAM, then to actual computation, and finally to physical AI implementation, cache coherence will be essential.”

There are significant trade-offs among these memory technologies. On-chip SRAM and caches provide fast access to frequently used data, but their capacity is limited by silicon area and cost. External DRAM offers greater capacity, while flash memory provides non-volatile storage for models, software, and other information that does not need to remain in active memory.

Moving information between these levels can reduce some capacity demands but increases transmission overhead and latency. Meanwhile, cache consistency focuses on keeping data copies in cache and memory synchronized, not merely transferring data between different storage technologies—but this also introduces additional overhead. Therefore, engineers must consider working set size, data locality, and access frequency when deciding how much memory to allocate at each level.

Automotive memory innovation is likely to play a significant role in how physical AI problems are solved.

Rob Fisher, Senior Director of Product Management at Imagination Technologies, said: “Currently, it is closely tied to our automotive product line, and most of the demand we see from robotics directly stems from automotive. Many automakers are now positioning their chips as both automotive and robotics platforms, so support for 4-bit, 6-bit, and 8-bit will significantly benefit memory bandwidth and memory footprint in embedded edge applications. This is less critical for data centers, but the impact differs: in data centers, you might run larger models by using narrower bit widths on word width; in automotive and robotics applications, you would instead leverage this to reduce memory footprint—you limit the model to only perform what it needs to do, thereby reducing the memory and bandwidth required to accommodate it.”

Lower-precision representations can reduce model storage and memory bandwidth requirements, making them particularly attractive for embedded AI. However, their benefits depend on whether the accelerator can efficiently support the chosen numerical format and whether quantization can maintain the precision required by the application. Formats such as 6-bit representation may also introduce packing or implementation overhead, limiting their benefits on hardware not specifically designed for them.

Flash memory also plays a crucial role in managing the memory of autonomous systems.

Sousa said: "DRAM is limited in terms of capacity, power consumption, and availability. Faster storage bandwidth can reduce load times, allowing the same DRAM budget to support larger models. These systems operate continuously on battery power and typically lack active cooling, so each data transfer consumes power and adds latency—meaning every component, from mechanical to electronic to software, must reduce power consumption and improve efficiency. Interfaces such as MIPI, UFS, LPDDR, and PCIe incorporate technologies that enable rapid burst data transfers, followed by extended low-power idle modes."

Faster non-volatile storage can reduce model loading latency and enable systems with limited DRAM to execute workloads whose total model footprint exceeds available active memory, when combined with appropriate model partitioning or weight streaming techniques. However, storage bandwidth alone does not increase the amount of data that can be accommodated in DRAM at any given moment.

The interfaces used to transmit this data also differ in their power management mechanisms and operational characteristics. Their efficiency depends on the specific specifications, implementation methods, transmission modes, and opportunities to enter low-power states.

If the market size is large enough, the memory requirements of physical AI may ultimately justify the development of new memory technologies or variants of existing ones.

Steven Woo, Distinguished Engineer and Distinguished Inventor at Rambus, said: “Memory is a critical technology that all physical AI will require. The difference is that some markets will only use existing memory, but there’s also the possibility that some markets will grow large enough to need their own memory type or a variant of existing memory. We’ve seen this with HBM. People were using other types of memory, and then the market suddenly expanded. If you go back to before LPDDR existed, phones were simply using DDR devices that had been bin-selected for low power consumption. Once a market reaches a certain scale, people say, ‘Well, let’s just create a dedicated device.’ Physical AI could grow into a large enough market that, if demand differences are significant enough, another type of memory could emerge.”

Whether this specialization occurs will depend on the scale and consistency of demand. Unlike data centers, physical AI applications have vastly different requirements in terms of cost, capacity, durability, bandwidth, temperature, and energy efficiency. This diversity may make some markets more inclined to adapt existing memory technologies, while others are better suited for purpose-built solutions.

Verify edge cases

Ensuring the system operates safely under all conditions, including rare and unexpected scenarios, is another major challenge. As physical AI developers target heavy machinery, medical devices, and other safety-critical technologies, addressing verification issues is especially important.

Balasubramanian said: “Before bringing physical AI into any production environment, establishing a validation loop is critical—I would say it’s the most important thing we need to do. For example, what happens during a large-scale power outage, an earthquake, or when external fiber optic cables are severed? How do you manage such scenarios at the edge? Keeping some form of computation, memory, and intelligence near the sensor will be crucial, and how people achieve this depends on identifying what matters most based on use case requirements. In an industrial setting, when do I need to stop the assembly line? When do I need to know to call back, trigger an alarm, or take other actions? These are all things that need to be handled within a single sensor.”

These scenarios illustrate that safety-critical functions cannot always rely on continuous cloud connectivity or centralized decision-making. Local processing may be necessary to detect failures, trigger protective actions, or bring devices into a defined safe state when communication or power systems fail.

However, verification and validation address different aspects of the problem. Verification ensures that the design meets established requirements, while validation checks whether the system performs appropriately in its intended operating environment. Safety assurance may also require hazard analysis, fault injection testing, common cause failure assessment, and evidence of compliance with applicable standards.

Therefore, the system must establish and verify trust around the relevant hardware and software stack, ensuring that security requirements align with each use case.

Owen Kent, co-founder of the assistive robotics startup ATDev, said: “When designing products for wheelchairs, as with any physical AI application, we have specific requirements. You must push the hardware to its limits within a quality management system and during third-party electromagnetic compatibility testing. Regardless of what hardware you use—whether custom or off-the-shelf—you must be able to verify its safety.”

For developers, this means not only evaluating whether components function correctly under normal operation, but also assessing how the complete system responds to sensor degradation, communication failures, unexpected inputs, and other foreseeable hazards. In many applications, passing component-level testing is a necessity, but it alone cannot prove system-level safety.

Conclusion

Physical AI's practical limitations span the entire pipeline from perception to action. Designers must align models and hardware with the task, move data within available power and thermal budgets, and verify machine behavior under changing conditions or connectivity. Increased computational power only truly helps when other parts of the system can support it.

Balasubramanian said: "Start with a standardized, low-power, and proven interface throughout the entire path—from sensors to processors to actuators. They provide designers with predictable behavior, ecosystem interoperability, and silicon-validated building blocks."

A relevant example is the MIPI Camera Security Framework, which combines the Camera Service Extensions (CSE) 2.0 specification with additional security specifications. CSE 2.0 supports data integrity protection for CSI-2 camera data, with optional encryption; the broader framework addresses authentication of imaging system components and secure session establishment.

He said, “It provides end-to-end security for camera data, covering integrity, authentication, and confidentiality from the image sensor to the SoC. It also offers designers flexible options to minimize computational and traffic overhead, power consumption, and latency, allowing security to be tailored to the application rather than designed for worst-case scenarios.”

This distinction is important because confidentiality is optional, and the level of protection depends on the security services and configurations implemented. The broader framework allows developers to balance protection against processing overhead, power consumption, and latency, rather than applying the same level of protection to every data stream.

Making autonomous machines more powerful is not enough. They must also operate reliably under real-world constraints. For physical AI, architectural choices that save memory, power, and bandwidth must also support predictable behavior, resilience, and the security required for large-scale deployment.

Data mobility extensions for physical AI: No one-size-fits-all solution.

Redefining the roles of edge and cloud AI: The local intelligence race is underway.

Brendan Heffernan

Disclaimer: The information on this page may have been obtained from third parties and does not necessarily reflect the views or opinions of KuCoin. This content is provided for general informational purposes only, without any representation or warranty of any kind, nor shall it be construed as financial or investment advice. KuCoin shall not be liable for any errors or omissions, or for any outcomes resulting from the use of this information. Investments in digital assets can be risky. Please carefully evaluate the risks of a product and your risk tolerance based on your own financial circumstances. For more information, please refer to our Terms of Use and Risk Disclosure.