Forem: Leonard Liao

Autonomous Robots and Edge AI

Leonard Liao — Wed, 29 Apr 2026 03:34:10 +0000

A few days ago, a small but important news story came out: a startup is trying to replace $100,000-per-day offshore ships with autonomous AI robots that can stay on-site and operate continuously.

At first glance, it sounds like just another robotics headline. But if you look closer, it highlights something much bigger - edge AI is no longer experimental, it’s becoming real infrastructure.

And this shift is happening faster than most people expected.

The key detail everyone misses

The offshore robotics example is just one signal.

Across the industry, we’re seeing the same pattern. Robotics systems are becoming autonomous, AI is moving out of data centers, and hardware is being optimized for local inference. A recent example is how generative AI robotics systems built by DeepX and Hyundai are pushing this transition even further toward real-world deployment.

This is what people now call physical AI - systems that don’t just analyze data, but actually act in the real world.

The interesting part of that story is not the robots themselves.

It’s how they work.

These systems don’t rely on constant cloud connectivity. They are designed to process data locally, make decisions in real time, and operate autonomously for long periods.

That’s a completely different architecture compared to traditional AI pipelines.

Instead of cloud → process → response, we now have device → process → action.

And that difference changes everything.

Why edge AI suddenly makes sense

This isn’t just hype. There are real engineering reasons behind it.

The biggest one is latency. Systems that operate in the physical world simply cannot wait for cloud responses. Robotics, industrial automation, and real-time monitoring all require immediate decisions, which is why processing is moving closer to where data is generated.

Bandwidth is another constraint. Streaming raw sensor data continuously is expensive and inefficient. Edge systems solve this by sending only processed results instead of raw input.

Then there’s reliability. If your system depends on connectivity, it can fail the moment the network becomes unstable. Autonomous edge systems don’t have that problem.

This is not an isolated case

The offshore robotics example is just one signal.

Across the industry, we’re seeing the same pattern. Robotics systems are becoming autonomous, AI is moving out of data centers, and hardware is being optimized for local inference.

This is what people now call physical AI - systems that don’t just analyze data, but actually act in the real world.

Hardware is the real driver here

None of this would be possible without changes in hardware.

Edge AI today is powered by specialized NPUs, efficient SoCs, and modular accelerator systems. Modern designs don’t rely on a single chip anymore. Instead, they combine control logic with dedicated inference hardware, allowing systems to scale performance without relying on the cloud.

That’s why modern architectures often look like this: a main processor for control, an AI accelerator for inference, and an optional cloud connection for training.

The shift toward distributed intelligence

What’s happening right now is a transition from centralized AI to distributed intelligence.

Instead of one powerful system doing everything, we now have thousands of smaller systems making decisions locally. This reduces latency, improves reliability, and makes systems far more scalable.

Inference is moving closer to the data source - into cameras, sensors, and machines - and that’s where the real change is happening.

If you want to understand the hardware side

Most discussions around edge AI stay at a high level. But the real difference comes from hardware choices - which chips are used, how they compare, and what trade-offs exist.

If you want a practical breakdown of that side, including real platforms and performance differences, this is worth checking edge AI hardware comparison and real-time analytics systems

It explains what actually powers modern edge systems and why modular AI architectures are becoming standard.

Where this is going next

The direction is already clear.

We’re moving toward systems that operate continuously, make decisions locally, and rely less on centralized infrastructure.

Not because it’s trendy, but because it’s the only way to make real-time systems actually work at scale.

AI Is Not Just “Killing Jobs” — It’s Rewiring the Tech Industry?

Leonard Liao — Sun, 26 Apr 2026 01:00:00 +0000

The recent analysis from Investopedia raises a question many people are quietly asking right now: is AI actually killing tech jobs, or just changing what those jobs look like?

The answer is not as dramatic as headlines suggest, but it is more important. What we are seeing is not a collapse of the tech industry. It is a structural shift inside it.

What’s Really Happening

In 2026, the tech sector has already seen over 100,000 layoffs. At the same time, companies like Microsoft, Meta, and Amazon are pouring massive investments into AI infrastructure.

At first, this looks contradictory. Why cut jobs while building more technology?

The explanation is simple. AI is not replacing entire companies. It is replacing specific layers of work inside them.

Tasks that used to require junior developers or support teams are now partially automated. Writing basic code, handling routine tickets, and processing structured data are no longer as labor-intensive as they were even two years ago.

This creates a visible pattern: fewer entry-level roles, but continued demand for high-level engineering and system design.

Disruption or Correction?

Some analysts describe this as disruption. Others call it normalization after the hiring boom during the pandemic.

Both views are valid.

Tech companies overexpanded between 2020 and 2022. Now they are correcting that. But AI is accelerating the correction. It gives companies a clear reason to streamline teams and rely more on automation.

There is also a growing belief that some companies are using AI as a narrative to justify layoffs that would have happened anyway. That does not mean AI is irrelevant. It means AI is both a real driver and a convenient explanation.

The Part Most People Miss: Infrastructure

Most discussions about AI and jobs focus only on software roles. But the bigger shift is happening at the infrastructure level.

As companies reduce repetitive human work, they increase investment in systems that can operate independently. This includes:

real-time data processing
on-device inference
edge computing

Instead of hiring more people to handle workflows, companies are building systems that can run those workflows automatically.

This is where hardware becomes critical.

Why Edge AI Matters Now

Cloud-based AI has limits. Latency, cost, and privacy concerns are pushing more workloads closer to the device and AI Chips, like Rockchip

That shift is driving demand for edge AI systems. These systems process data locally, respond in real time, and reduce reliance on centralized servers.

This is already happening in areas like robotics, smart cameras, and industrial automation.

Platforms built on chips like RK3588 are part of this transition. They allow developers to run AI models directly on devices instead of sending everything to the cloud.

If you want a clear technical breakdown of how this works in practice, this article explains the fundamentals of edge AI for real-time analytics on Rockchip platforms
and why local processing is becoming the default approach.

Fewer Jobs, More Output

Every major technological shift has reduced the need for repetitive human work.

AI is continuing that pattern, but this time inside the tech industry itself.

Teams are getting smaller, but more productive. A few engineers with the right tools can now do the work that previously required entire departments.

This does not mean jobs disappear completely. It means the nature of those jobs changes.

What Comes Next

In the short term, we will likely see:

fewer entry-level positions
slower hiring across big tech
more automation in routine workflows

In the long term, the focus shifts toward:

AI system design
hardware-aware development
edge and real-time computing

The demand is not going away. It is moving.

Where RK182X Fits In

Leonard Liao — Sat, 25 Apr 2026 02:06:33 +0000

The conversation around robotics has changed a lot in the past few years. What used to be mostly about movement and control is now heavily focused on intelligence. Robots are no longer just executing predefined actions — they are expected to perceive, understand, and react in real time.

This shift puts serious pressure on hardware.

The Real Bottleneck in Robotics Today

If you look at modern robotics systems, most of the complexity comes from combining multiple workloads at once:

motion control
vision processing
AI inference
real-time decision making

Trying to run all of this on a single chip often leads to compromises. Either you sacrifice latency, or you limit the size of AI models you can run.

This is exactly why hybrid architectures are becoming more common.

RK3588 + RK182X: Splitting the Workload

A practical approach is to separate responsibilities between chips.

The RK3588 already handles:

system control
sensor input
video pipelines
general-purpose compute

But when it comes to heavier AI workloads — especially large models — it makes sense to offload them.

That’s where RK182X comes in.

Instead of overloading the main SoC, the AI co-processor handles inference independently. This avoids resource contention and keeps real-time systems stable.

Why This Matters in Real Deployments

This architecture is not just theoretical.

In real robotics systems, timing matters a lot. Even small delays in decision-making can affect navigation, interaction, or safety.

By separating AI inference from control logic:

latency becomes more predictable
system stability improves
scaling AI models becomes easier

This is especially important for applications like:

humanoid robots
industrial automation
autonomous inspection systems

A Broader Industry Trend

This shift toward dedicated AI accelerators is not unique to Rockchip.

According to recent industry analysis by McKinsey & Company, companies are increasingly investing in edge AI infrastructure that enables real-time processing directly on devices, rather than relying on cloud-based systems.

This trend is driven by:

lower latency requirements
privacy concerns
reliability in offline environments

Where RK182X Stands Out

As the industry moves forward, attention is also shifting toward next-generation AI chips. For example, there is already growing discussion around upcoming architectures like RK3688 and what they could bring to edge AI systems. A recent overview of Rockchip’s next-gen RK3688 AI SoC highlights how future designs may push performance even further.
What makes RK182X interesting is how it fits into a complete platform.

It is not just an isolated chip. It is designed to work alongside RK3588 as part of a full robotics stack that includes:

optimized vision pipelines
audio interaction systems
AI model libraries
ROS2-based frameworks

If you want a deeper technical breakdown of how this platform is structured and what capabilities it includes, this detailed overview of the RK182X AI processor for robotics covers the full stack and real deployment scenarios.

From Demos to Real Products

One of the biggest changes in robotics right now is the transition from prototypes to production systems.

We are seeing:

robots deployed in factories
AI-powered logistics systems
autonomous inspection tools

These are not experiments anymore.

They are real systems running in real environments, which means hardware decisions matter more than ever.

Final Thoughts

The future of robotics is not just about more powerful chips — it is about smarter system design.

Splitting workloads between general-purpose SoCs and dedicated AI processors is one of the most practical ways to scale performance without sacrificing stability.

RK182X is a good example of this approach.

It reflects a broader industry direction:
move intelligence closer to the device, reduce latency, and make systems more reliable in real-world conditions.

DEEPX and Hyundai Are Building Generative AI Robots

Leonard Liao — Wed, 22 Apr 2026 02:43:33 +0000

What It Actually Means for Edge AI

A recent Reuters report highlights a major shift in edge AI: real robotics platforms are now being designed specifically for generative AI workloads.

This is not just another partnership announcement. It shows how edge AI hardware, robotics, and generative models are starting to converge into real products.

What Happened

South Korean AI chip startup DEEPX is expanding its partnership with Hyundai Motor Group to build a new computing platform for robots powered by generative AI.

The key element is DEEPX’s upcoming DX-M2 chip, a second-generation low-power NPU designed for on-device AI.

Unlike cloud-based AI systems, these chips run models directly on robots — meaning:

no dependency on cloud latency
lower power consumption
real-time decision making

The chips will be manufactured using Samsung’s 2nm process, with mass production planned for next year.

Why This Is Important

This news confirms a major trend: generative AI is moving to the edge.

DEEPX is not building general-purpose GPUs. It focuses on NPUs optimized for:

robotics
autonomous systems
industrial AI

According to the company, its current chips are up to 20× more power-efficient and cheaper than NVIDIA Jetson Orin.

That is a big deal for robotics, because power and heat are major constraints, especially for humanoid robots.

Generative AI Inside Robots

One of the most interesting parts of the report is how DEEPX describes the next generation of chips.

They are designed specifically for generative AI workloads, meaning robots can:

learn from experience
adapt to new environments
improve behavior over time

This is closer to how large language models work, but applied to physical systems.

Instead of fixed logic, robots become systems that can evolve.

Hyundai’s Strategy: Scale Robotics Production

Hyundai is not experimenting — it is scaling.

The company plans to produce up to 30,000 robots per year by 2028, including humanoid platforms.

This partnership with DEEPX is part of a broader strategy:

build an ecosystem of on-device AI
reduce reliance on external compute
control the full robotics stack

This is similar to what Apple did with its own silicon — but now for robotics.

Why Low-Power NPUs Matter

Traditional AI hardware like GPUs is powerful but inefficient for edge systems.

Robots need:

low latency
low power consumption
stable thermal behavior

DEEPX specifically targets these constraints.

For example, the company notes that lower power consumption helps prevent overheating in humanoid robots — a real problem for current designs. ([Investing.com][2])

This is where specialized AI accelerators outperform general-purpose hardware.

How This Relates to Existing Edge AI Hardware

If you look at the current edge AI ecosystem, there are already several approaches:

NVIDIA Jetson (high performance, higher power)
Hailo accelerators (efficient NPU design)
emerging players like DEEPX

Also, a good breakdown of DeepX Architectures

What DEEPX is doing is pushing this space further toward generative AI + robotics, not just inference.

Bigger Picture

This news confirms several important shifts in edge AI:

AI is moving from cloud → on-device
robotics is becoming a primary use case
NPUs are replacing GPUs in many edge scenarios
generative AI is entering real-world systems

Most importantly, this is no longer theoretical.

These systems are being designed for mass production.

Conclusion

The DEEPX + Hyundai partnership is a clear signal that edge AI is entering a new phase.

This is not about benchmarks or demos.

It is about building real robots powered by on-device generative AI.

And once this model works, it will spread fast across:

factories
logistics
autonomous machines

Edge AI is becoming physical.

Edge AI News – April 2026

Leonard Liao — Wed, 22 Apr 2026 02:33:50 +0000

Edge AI News – April 2026: Real Developments in Robotics, Chips, and Industrial AI

April 2026 was not about hype. It was about real deployments, real hardware, and real partnerships shaping edge AI.

NVIDIA and Cadence Push AI Training for Robotics

One of the most important news this month came from a partnership between NVIDIA and Cadence.

They are working on combining AI with physics-based simulation to train robots faster and more safely. Instead of training robots only in the real world, companies can now generate synthetic data using accurate simulations.

This reduces training time and cost significantly, especially for complex robotic tasks. (Reuters)

This is a key step toward scalable robotics development.

Hyundai + DEEPX: Low-Power Edge AI Chips for Robots

Another major move comes from South Korea.

AI chip startup DEEPX is working with Hyundai to build robots powered by low-power edge AI chips. These chips are designed for real-world deployment, not cloud inference.

Key details:

Focus on robotics, factories, autonomous systems
Up to 20× better efficiency vs traditional solutions
Mass production planned using advanced semiconductor nodes

Hyundai plans to scale robot production to tens of thousands per year, which shows this is not experimental anymore. (Reuters)

Qualcomm Enters Edge AI SBC Market (Jetson Alternative)

Qualcomm is directly attacking NVIDIA’s position in edge AI hardware.

They introduced a new single-board computer for robotics with:

~40 TOPS AI performance
ARM architecture
Price under $300

This is important because it lowers the barrier to entry for developers and startups building edge AI systems. (Futurum)

This could seriously impact the Jetson ecosystem in the next 1–2 years.

Hannover Messe 2026: Edge AI Is Already in Factories

At Hannover Messe 2026, edge AI was not theoretical — it was deployed.

Key highlights:

AI-driven robotics in manufacturing
real-time defect detection using computer vision
digital twins + edge AI for production optimization

Manufacturers are now using edge AI to detect defects instantly and react in real time. (Lenovo StoryHub)

NVIDIA also demonstrated full AI-powered factory systems, including robotics, simulation, and autonomous workflows. (NVIDIA Blog)

This confirms that edge AI is already production-ready.

Sensor Fusion: Radar + Vision for Robotics

A practical development this month is sensor fusion architecture for robotics.

Companies like Texas Instruments and Lattice introduced systems combining:

mmWave radar
cameras
FPGA-based synchronization

The goal is low-latency perception pipelines for robots operating in real environments. (Business Wire)

This is critical because real-world AI requires reliable perception, not just raw compute.

Safety-Critical Edge AI Platforms (QNX + NVIDIA)

Another important step is the move into regulated industries.

QNX and NVIDIA announced a unified platform combining:

real-time OS
AI compute (NVIDIA IGX Thor)
functional safety stack

This is designed for:

robotics
medical devices
industrial systems

The key idea: edge AI must be deterministic and reliable, not just fast. (ACCESS Newswire)

NPU Explosion: Edge AI Everywhere

Outside robotics, edge AI is expanding into consumer devices.

Modern devices are already reaching:

~100 TOPS in smartphones
projected 400–1000 TOPS combined per user by 2030

This means edge AI is becoming distributed across devices, not centralized in the cloud. (TechRadar)

Conclusion

April 2026 shows a clear shift in edge AI:

Real partnerships (NVIDIA + Cadence, Hyundai + DEEPX)
Real hardware competition (Qualcomm vs NVIDIA)
Real deployments (factories, robotics)
Real constraints (latency, safety, power)

Edge AI is no longer a concept.

It is infrastructure.

RK3588 vs Jetson Orin Nano: Real-World comparison

Leonard Liao — Sat, 11 Apr 2026 02:04:32 +0000

When people talk about edge AI hardware in 2026, two names come up again and again: RK3588 and Jetson Orin Nano. Both are widely used for building compact AI systems, robotics projects, and embedded computer vision setups. On paper, they look very different. But in real-world use, the gap is not always what you expect.

Let’s break it down in simple terms and look at how these two platforms actually behave outside of benchmarks.

CPU and general performance

RK3588 uses an 8-core CPU with a mix of high-performance Cortex-A76 cores and power-efficient Cortex-A55 cores. This makes it very flexible. It can handle heavy workloads, but it can also run lightweight tasks without wasting power.

Jetson Orin Nano is more focused on AI acceleration. Its CPU is not as strong in general-purpose tasks. In real-world scenarios like running a full Linux system with multiple services, RK3588 often feels smoother.

If your project includes not just AI, but also backend logic, UI, or multitasking, RK3588 has a clear advantage.

For a deeper breakdown of the chip architecture and real capabilities, you can check this detailed RK3588 overview.

AI performance and acceleration

This is where Jetson usually wins on paper. NVIDIA provides a strong AI ecosystem with CUDA, TensorRT, and optimized libraries.

However, RK3588 is not weak. It includes an NPU capable of around 6 TOPS. In real-world applications like object detection, face recognition, or simple tracking, it performs well enough for most edge use cases.

The key difference is not just raw power, but ecosystem. Jetson has better software tools. RK3588 requires more manual setup, but gives more flexibility.

In many production environments, especially cost-sensitive ones, RK3588 is often “good enough” while being much cheaper.

Power consumption and efficiency

Power efficiency is one of the biggest reasons people choose RK3588.

Jetson Orin Nano can draw more power under load. RK3588 systems are usually easier to run passively cooled or with minimal thermal design.

For edge deployments, kiosks, or always-on devices, this matters more than raw AI performance.

Real-world use cases

In practice, these chips are used in slightly different ways.

Jetson Orin Nano is common in:

robotics research
autonomous systems
advanced AI prototyping

RK3588 is widely used in:

smart displays
retail analytics
industrial control systems
media + AI hybrid devices

If your project needs both video processing and AI, RK3588 often feels more balanced.

You can also see a real hardware implementation based on this chip here as an RK3588-based device example

Cost and scalability

This is where RK3588 becomes very attractive.

Jetson boards are more expensive, and scaling to multiple devices can quickly increase costs.

RK3588-based systems are much cheaper, and there are many vendors offering different form factors. This makes it easier to scale projects or build custom hardware.

Ecosystem and development experience

NVIDIA has a clear advantage in developer experience. Their documentation, SDKs, and community are very strong.

RK3588 is improving, but still requires more low-level work. You may need to configure drivers, optimize models manually, or adapt frameworks.

That said, for teams with embedded experience, this is not a problem.

Final thoughts

There is no universal winner.

Jetson Orin Nano is better if you want:

strong AI ecosystem
faster setup
advanced deep learning workflows

RK3588 is better if you want:

lower cost
better overall system performance
flexible edge deployment

In real-world projects, the choice often depends on constraints, not just specs.

If you are building something practical and cost-sensitive, RK3588 is often the smarter option. If you are focused on AI-first development with minimal setup, Jetson may be easier.

For additional technical context on edge AI hardware trends, this overview from NVIDIA is also useful:

Choosing the Right Desk Occupancy Sensor: A Practical Guide

Leonard Liao — Sun, 08 Feb 2026 03:33:13 +0000

Introduction: Why Desk Occupancy Detection Matters Today

The evolution of hybrid work has fundamentally changed how offices operate. Organizations are no longer focused solely on how many employees they have, but rather on how physical workspace is actually used throughout the day. Empty desks in fully leased offices represent wasted cost, while overcrowded zones reduce productivity and employee comfort.

This is where the desk occupancy sensor becomes a critical component of modern workplace strategy. By identifying whether a specific workstation is in use, these sensors enable real-time space analytics, automated desk-booking enforcement, and long-term planning based on real behavioral data rather than assumptions.

However, detecting occupancy at the desk level is significantly more complex than detecting motion in a meeting room. Employees may sit still for long periods, objects may be left on chairs, and privacy expectations are much higher. Selecting the correct sensing technology, therefore, requires careful evaluation of accuracy, deployment constraints, and data governance.

Understanding What Desk Occupied Really Means

Before choosing any hardware, organizations must define what occupancy actually represents in their workflow.

For some environments, occupancy simply means a person is physically present at the workstation. In others, it refers specifically to a chair being used, regardless of activity. More advanced interpretations may include active work detection, distinguishing between a briefly abandoned desk and one that has been vacated for good.

Effective systems rarely rely on a single instant signal. Instead, they incorporate timing logic, confidence scoring, and hold periods that prevent desks from being marked vacant too quickly. This subtle software layer is often just as important as the sensing technology itself.

Core Desk Occupancy Sensor Technologies Explained

Multiple sensing approaches are used across smart office deployments, each balancing cost, precision, and privacy in different ways.

Motion-based infrared sensing remains one of the simplest and most affordable options. These sensors detect changes in heat patterns caused by human movement. While suitable for broad activity detection, they can struggle in desk scenarios where a user remains still for extended periods.

Ultrasonic sensing improves sensitivity by measuring reflections of high-frequency sound waves. It can capture subtle motion but is vulnerable to environmental interference such as airflow, HVAC noise, or nearby sensors operating in the same frequency range.

Millimeter-wave radar represents a more advanced presence-detection method. By analyzing reflected radio waves, it can identify micro-movements such as breathing or slight posture changes. This enables reliable detection even when a person is motionless, making it particularly effective for desk-level monitoring in real offices.

Pressure-based sensing takes a different approach by measuring physical weight on a chair or seat surface. This method provides clear confirmation of seated presence, though it may misinterpret heavy objects as occupants and can introduce installation complexity during retrofitting of existing furniture.

Capacitive proximity sensing detects disturbances in an electrical field near the desk. While discreet and inexpensive, its reliability depends heavily on materials, mounting position, and surrounding electromagnetic conditions.

Device-based detection infers occupancy from smartphones or laptops connected via Bluetooth or Wi-Fi. Although convenient in managed IT environments, it measures device presence rather than human presence, raising both accuracy and privacy considerations.

Vision and depth-based sensing technologies provide the richest spatial awareness, but workplace privacy expectations and regulatory constraints often limit their adoption. Where used, they typically rely on anonymized, on-device processing that outputs only occupancy states rather than images.

Key Factors When Selecting a Desk Occupancy Sensor

Choosing the right desk occupancy sensing solution is ultimately a balance of priorities.

Accuracy is essential when occupancy data drives automation, such as releasing unused reservations or optimizing cleaning schedules. Technologies capable of detecting still presence-like radar or pressure sensing-tend to perform best in these scenarios.

Privacy has become equally critical. Employees are far more comfortable with systems that output anonymous occupancy states rather than identifiable behavioral data. Even non-visual tracking methods can raise concerns if they associate devices or identities with desks.

Deployment complexity also shapes real-world decisions. Battery-powered sensors reduce installation effort but require maintenance cycles, while wired solutions increase upfront costs but simplify long-term operation. Office density, partitions, and furniture layouts further influence which sensing method performs reliably.

Placement, Signal Fusion, and Real-World Performance

Sensor placement dramatically affects detection quality. Mounting beneath the desk can discreetly capture torso presence, while above-desk placement improves line-of-sight detection but risks obstruction. Chair-integrated sensing excels for seated work yet may require complementary methods for sit-stand environments.

Because no single technology is perfect, many advanced deployments rely on sensor fusion-combining multiple signals to improve confidence and reduce false readings. Intelligent firmware, adaptive thresholds, and time-based logic transform raw sensor input into meaningful occupancy insight.

The Future of Smart Desk Sensing in Hybrid Offices

As workplaces continue shifting toward flexible usage models, desk occupancy sensing is moving beyond simple presence detection. Integration with booking platforms, environmental controls, and analytics dashboards is turning occupancy data into a core operational signal for smart buildings.

Advances in low-power radar, edge processing, and privacy-preserving AI are expected to further improve reliability while maintaining employee trust. Over time, desk occupancy sensing will likely become a standard infrastructure layer, similar to Wi-Fi or lighting control, quietly enabling more efficient, responsive workplaces.

Conclusion

Selecting a desk occupancy sensor is not merely a hardware decision but a broader system design challenge. True success depends on aligning sensing technology with organizational goals, privacy expectations, and deployment realities.

Basic motion detection may be sufficient for trend analysis, but accurate real-time understanding of desk usage typically requires technologies capable of sensing still presence or direct physical occupancy. When thoughtfully implemented, desk occupancy sensing provides measurable benefits in cost efficiency, workplace experience, and long-term space strategy-making it a foundational element of the modern smart office.

RK3588S2: A Powerful ARM-Based SoC

Leonard Liao — Thu, 22 Jan 2026 07:01:55 +0000

The RK3588S2 is Rockchip’s latest addition to its high-performance ARM-based processors. Designed to address the growing needs of PC, edge computing, and digital multimedia applications, the RK3588S2 integrates powerful processing cores, a versatile GPU, advanced video codecs, and AI acceleration. With this combination, the RK3588S2 is poised to power next-generation devices that require strong compute performance, efficient multimedia processing, and robust edge AI inference.

Key Highlights

1. Microprocessor
At the heart of the Rockchip RK3588S2 is a heterogeneous octa-core CPU cluster:

1) Quad-core Cortex-A76 MPcore processors for high-performance workloads.

2) Quad-core Cortex-A55 MPcore processors for power-efficient tasks.

This big.LITTLE setup balances performance and efficiency, while the DynamIQ Shared Unit (DSU) allows flexible workload sharing across cores. Each Cortex-A76 has 64KB L1 cache (instruction + data) and a 512KB L2 cache, whereas each Cortex-A55 has 32KB L1 cache and a 128KB L2 cache. The chip also offers a shared 3MB L3 cache.

2. Graphics and Multimedia

The RK3588S2 integrates a powerful embedded 3D GPU, supporting:

1) OpenGL ES 1.1/2.0/3.2

2) Vulkan 1.2

It also comes with a specialized 2D hardware engine with MMU to optimize display performance.

For video processing, the SoC supports:

1) 8K@60fps decoding for H.265 and VP9

2) 8K@30fps decoding for H.264

3) 4K@60fps decoding for AV1

4) H.264 and H.265 encoding up to 8K@30fps

Additionally, it features a JPEG encoder and decoder along with a variety of specialized pre- and post-processors for image enhancement.

3. Advanced ISP (Image Signal Processor)

The RK3588S2 features a hardware-based ISP capable of processing up to 48-megapixel images. It includes accelerators for advanced imaging algorithms, such as:
1) HDR
2) 3A (Auto-Exposure, Auto-Focus, Auto-White Balance)
3) LSC, 3DNR, 2DNR
4) Sharpening, dehaze, fisheye correction, gamma correction

This makes it especially ideal for camera-based edge AI applications, intelligent surveillance, and computer vision systems.

4. AI Acceleration

A major feature of the RK3588S2 is its built-in NPU (Neural Processing Unit), which supports hybrid operations including INT4, INT8, INT16, and FP16. Offering up to 6 TOPS of processing power, it provides robust performance for edge AI applications. Moreover, it is compatible with widely used machine learning frameworks such as TensorFlow, PyTorch, MXNet, and Caffe, facilitating easy transfer and deployment of AI models.

5. Memory and Interfaces

The RK3588S2 accommodates high-performance memory interfaces such as LPDDR4, LPDDR4x, and LPDDR5, ensuring sufficient bandwidth for demanding tasks. Its versatility allows it to be used across various applications, from embedded systems to lightweight edge servers.

Applications

The RK3588S2 is extremely versatile and can be used in a wide range of applications:
2) Edge Computing Gateways: Handling AI inference, image recognition, and local decision-making.

3) Digital Signage and Displays: Powering 8K playback and smooth rendering.

4) Smart Cameras: Utilizing its advanced ISP and AI processing for analytics.

5) Mobile Internet Devices: Offering strong multimedia and connectivity features.

AIoT Systems: Enabling efficient real-time AI at the edge.

Conclusion

The Rockchip RK3588S2 is a powerful SoC that features high-performance ARM cores, an advanced GPU, strong multimedia codecs, a reliable ISP, and an integrated NPU. It supports 8K multimedia and offers AI acceleration at the edge, making it ideal for developers building next-generation smart devices, embedded systems, and AI-driven applications. For those looking for a good mix of performance, efficiency, and flexibility, the RK3588S2 is among the best ARM-based options available today.

Rockchip RK3588 Boards 2026

Leonard Liao — Tue, 20 Jan 2026 03:35:29 +0000

This comprehensive overview examines the Rockchip RK3588 boards ecosystem in 2026, compares popular options, and guides you in choosing the right one for projects like edge computing, robotics, embedded AI, and media servers.

Rockchip RK3588

The Rockchip RK3588 is an 8-core Arm System on Chip (SoC) built for high-performance embedded computing. It includes:

Features include 4 Cortex-A76 performance cores and 4 Cortex-A55 efficiency cores, an ARM Mali-G610 GPU for advanced graphics and multimedia, a dedicated 6 TOPS NPU for AI tasks, support for a wide range of video codecs and displays, and extensive connectivity options such as PCIe, USB3, Ethernet, and multi-camera interfaces.

Fabricated using an 8 nm process node, the RK3588 offers robust compute, graphics, and AI capabilities with efficient power use. It is expected to be highly popular in 2026 for SBCs, industrial edge devices, and AIoT products.

Comparison Table

Professionals often choose RK3588 boards for self-hosted services like media servers, such as Jellyfin, containers, or lightweight edge servers, due to their efficient power usage. Equipped with robust connectivity and a wide range of I/O options, these boards are ideal for robotics controllers, automation gateways, and custom embedded systems where space and power constraints are critical.

The ecosystem’s expansion emphasizes how RK3588 keeps bridging the divide between hobbyist boards and industrial solutions, offering desktop-level performance alongside scalable connectivity and extensive software support. For customers aiming to deploy RK3588 boards on a large scale or with tailored hardware needs, professional support and design services are accessible to help speed up development and deployment.

Conclusion

So, here are the best RK3588 single-board computers (SBCs):
1) Kiwi Pi 5B
2）Kiwi Pi 5 Pro / Ultra
3）Orange Pi 5 Plus / Pro
4）Radxa Rock 5B

By 2025, Rockchip RK3588 boards continue to be an attractive option for developers who need a combination of high performance, AI features, and multimedia capabilities in a small size. From the Kiwi Pi 5B to the more advanced Kiwi Pi 5 Pro/Ultra, as well as competing options from Orange Pi and Radxa, these SBCs offer strong platforms for embedded systems, AI edge applications, media servers, and other projects.

The Silent Sentinel: How Occupancy Sensors Are Quietly Revolutionizing Our Spaces

Leonard Liao — Fri, 16 Jan 2026 09:38:04 +0000

In the ever-evolving landscape of smart buildings and sustainable design, one technology operates with such subtlety that its profound impact often goes unnoticed. The occupancy sensor—a small, unassuming device—is fundamentally transforming how we interact with and manage our environments, delivering remarkable gains in efficiency, comfort, and cost savings.

Beyond the Simple Light Switch

At its core, an occupancy sensor is an electronic device that detects the presence or absence of people within a monitored space. While most famously associated with automatically turning lights on when you enter a room and off when you leave, its applications now extend far beyond illumination.

These sensors employ various technologies to perform their task:

Passive Infrared (PIR): The most common type, it detects heat emitted by moving human bodies. It's cost-effective and ideal for smaller, enclosed spaces like private offices or restrooms.

Ultrasonic: Emits high-frequency sound waves and measures their reflection. It can detect subtle movements, like typing at a desk, making it perfect for spaces with obstructions or where people remain still for long periods.

Microwave: Similar to ultrasonic but uses radio waves. It offers a wider coverage area and can sense through non-metallic materials, suited for larger, open-plan areas.

Dual-Technology (DT): Combines PIR and another technology (often ultrasonic) to reduce false triggers. It requires both sensors to be triggered for an "occupied" state, maximizing accuracy.

Image Processing & Camera-Based: Advanced systems using low-resolution cameras or people-counting algorithms to not only detect presence but also track space utilization patterns.

The Triple Bottom Line: Why Occupancy Sensors Are Essential

The adoption of these intelligent sentinels is driven by a powerful convergence of benefits:

1. Energy Efficiency & Sustainability
This is the most celebrated advantage. By ensuring lights, HVAC, and plug loads are only active when needed, waste is dramatically reduced. The U.S. Department of Energy notes that lighting alone accounts for about 17% of a commercial building's electricity use—a significant portion of which is wasted in empty rooms. Automated control via occupancy sensors can yield lighting energy savings of 30% or more. In an era of climate consciousness and rising energy costs, this is a compelling proposition.

2. Operational Cost Reduction
Energy savings translate directly into lower utility bills. Furthermore, reduced runtime extends the lifespan of lighting fixtures, HVAC components, and other equipment, slashing maintenance and replacement costs. For large-scale commercial real estate portfolios, these savings compound into substantial financial returns.

3. Enhanced Convenience, Safety, and Data Insights
The "hands-free" operation adds a layer of hygiene and convenience, particularly in restrooms, storage areas, or when carrying items. In corridors and stairwells, they enhance security by ensuring areas are well-lit upon approach. Modern, network-connected sensors provide a treasure trove of space utilization data. Facility managers can see which conference rooms are actually used, how often workstations are occupied, and optimize cleaning schedules or space allocations based on real data rather than guesswork.

Intelligent Integration: The Heart of the Smart Building

Today's occupancy sensors are not isolated devices. They are integral nodes in the Internet of Things (IoT) ecosystem for smart buildings. Integrated with Building Management Systems (BMS), they enable holistic control:

HVAC Optimization: Ventilation and temperature can be zoned and adjusted based on real-time occupancy, providing comfort where needed and saving energy where not.

Space Management: Data feeds into workplace apps, allowing employees to find and book available rooms or desks dynamically.

Predictive Maintenance: Systems can alert facilities teams when areas see abnormal usage patterns or when sensor health is declining.

Considerations and Best Practices

Successful deployment requires careful planning:

Placement is Key: Sensors must be located to cover the entire zone without obstructions, avoiding false triggers from vents, windows, or HVAC units.

Technology Selection: Choose the right sensor type for the space. A PIR sensor might fail in a restroom stall, while ultrasonic is ideal for a focused library cubicle.

Timeout Settings: Adjust the delay period before a space is considered "vacant" to balance energy savings with user annoyance (no one wants the lights to go out while reading).

Privacy: For camera-based or more advanced systems, clear communication about data collection and anonymization is essential to maintain trust.

The Future: More Adaptive and Predictive

The future of occupancy sensing lies in increased intelligence and granularity. We are moving towards systems that don't just detect if a space is occupied, but how many people are present and even what they are doing. This allows for finer-tuned environmental control. Furthermore, AI and machine learning will enable predictive occupancy, anticipating room usage based on historical patterns, meetings schedules, and even real-time location data (with user consent), preparing spaces just before they are needed.

Conclusion

The humble occupancy sensor has evolved from a simple light-switch helper to a critical data-gathering tool for intelligent space management. It stands as a perfect example of how subtle, automated technology can create outsized benefits—making our buildings greener, our operations leaner, and our environments more responsive. In the quest for smarter, more sustainable, and human-centric spaces, the silent sentinel is on constant duty, ensuring that energy and resources are devoted precisely where and when they matter most.

Arm Mali-G57 MC2 GPU: Mid-Range Graphics Power

Leonard Liao — Fri, 26 Dec 2025 02:26:43 +0000

The ARM Mali-G57 MC2 is a mid-range graphics processing unit (GPU) from ARM Holdings that brings modern graphics performance and power efficiency to a wide range of mobile and embedded devices. Positioned between entry-level and high-end solutions, the G57 MC2 bridges performance and efficiency, making it a compelling choice for mainstream smartphones, tablets, smart TVs, and other graphics-enabled systems.

This article explores the architecture, key features, performance characteristics, real-world use cases, and comparative position of the Mali-G57 MC2 in today’s GPU landscape.

What Is the Mali-G57 MC2 GPU?

The Mali-G57 MC2 is part of ARM’s Mali-G57 series, based on the Valhall architecture—a design that emphasizes improved performance, better energy efficiency, and strong support for modern graphics APIs like Vulkan and OpenGL ES. The MC2 suffix indicates that this GPU configuration includes two cores (clusters), offering balanced graphics performance for devices that need more than basic UI rendering but don’t require flagship GPU power.

Architecture & Technology: Valhall at Its Core

The Valhall architecture represents a generational improvement over previous Mali designs:

Enhanced performance density: Compared with earlier Mali GPUs like the G52, the G57 delivers around 30% higher performance density and double the texturing performance, enabling smoother 3D graphics and better support for high-resolution content.

Parallel compute design: Each core has multiple ALUs (Arithmetic Logic Units), enabling efficient parallel processing for graphics rendering and compute tasks.

API support: The G57 MC2 supports Vulkan, OpenGL ES, and OpenCL, allowing developers to build games and applications that leverage modern graphics pipelines.

Despite its mid-tier positioning, this GPU also integrates features like advanced shading, HDR support, and texture compression — helping developers extract more visual fidelity from mid-range hardware.

Performance Overview

Dual-Core (MC2) Configuration

The MC2 configuration (two cores) strikes a balance between performance and efficiency. Benchmarks like Geekbench Vulkan scores (e.g., ~1298 on some devices) illustrate its mid-range nature — capable of handling everyday graphics workloads and casual gaming smoothly, but obviously not competing with flagship GPUs.

In independent performance tables and specs, typical Mali-G57 MC2 configurations deliver modest shader counts and moderate FP32/FP16 throughput, suitable for 1080p or lower-resolution rendering at playable frame rates on mainstream titles.

Power Efficiency & Battery Life

One of the Mali-G57’s biggest strengths is power efficiency — a critical factor in battery-constrained devices like phones and tablets:

Designed for mobile use cases, the Valhall architecture delivers improved performance per watt compared to earlier Mali generations.

Features like clock and power gating help decrease energy consumption during periods of low workload.

This leads to longer battery life during multimedia playback and light gaming — a key advantage in the mainstream device market.

Final

The Arm Mali-G57 MC2 offers a strong solution for mainstream mobile and embedded graphics:

✔️ Balanced performance for everyday use
✔️ Good energy efficiency for mobile devices
✔️ Modern API support and architectural improvements
✔️ Scalable integration in SoCs targeting broad hardware tiers

While not ideal for flagship-level gaming or intensive professional tasks, the G57 MC2 delivers strong graphics performance for mid-range devices, helping manufacturers deliver responsive UIs and enjoyable media experiences without excessive power consumption.

In summary: The Mali-G57 MC2 is a modern, efficient, and capable mid-range GPU that hits the sweet spot for mainstream mobile graphics — perfect for daily users seeking balanced performance and battery life.

Source: Mali-G57 MC2 GPU: Practical Overview

Arm Immortalis-G925: The Next-Gen Flagship Mobile GPU

Leonard Liao — Fri, 26 Dec 2025 02:14:20 +0000

The Arm Immortalis-G925 represents a major evolution in mobile GPU technology, ushering in a new era of high-performance graphics and artificial intelligence (AI) performance for flagship devices. Built on Arm’s 5th-generation GPU architecture, Immortalis-G925 is Arm’s most powerful and efficient GPU to date, pushing advancements in gaming visuals, ray tracing, and AI workloads while simultaneously reducing power consumption compared to prior generations.

What Is Immortalis-G925?

Arm’s Immortalis line sits at the top tier of its GPU offerings—above the Mali family—with hardware support for advanced features such as ray tracing. The Immortalis-G925 is designed specifically for flagship mobile SoCs, targeting premium smartphones, tablets, and other high-performance, power-sensitive applications

Key architectural enhancements include:

Fragment prepass – reduces redundant rendering work to improve efficiency.

Doubled tiler and shift-convert throughput – accelerates geometry and shading workloads.

Improved command stream frontend – optimizes job dispatch to shader cores.

Enhanced ray tracing performance – facilitates realistic lighting and effects.

Scalable configurations – supports up to 24 GPU cores to suit diverse performance points.

These architectural changes help balance peak performance and sustained efficiency, a key challenge in thermal-limited smartphone environments. Unlike solutions that chase peak scores only to throttle quickly under load, G925’s design focuses on delivering consistent real-world gaming performance over longer sessions.

Performance Gains: Gaming, Efficiency, and Beyond

Arm’s internal tests and industry analyses indicate that Immortalis-G925 delivers significant performance uplifts over its predecessor, the Immortalis-G720:

Graphics & Gaming

1) ~37 % higher overall graphics performance compared to Immortalis-G720.

2) Up to ~46 % average uplift in popular mobile games.

3) Standout game performance improvements include Genshin Impact (~49 %) and Roblox (~46 %).

4) 52 % better ray tracing performance with intricate scenes, enabling more realistic shadows, reflections, and lighting.

This makes Immortalis-G925 particularly attractive for mobile AAA gaming, where high frame rates and advanced visual effects are increasingly expected on flagship phones.

Efficiency & Power

Despite the performance gains, Immortalis-G925 uses ~30 % less power than the previous generation when delivering comparable graphics performance.

This efficiency boost is crucial for maintaining battery life and mitigating thermal throttling in handheld devices.

AI & Machine Learning Capabilities

Beyond raw graphics, Immortalis-G925 also enhances AI and machine learning performance. Arm’s GPU architecture can accelerate various AI inference tasks, such as image classification, segmentation, object detection, and natural language processing—helpful for on-device AI features in mobile apps

In benchmark tests, the GPU demonstrated:

~36 % faster AI inference performance vs. the previous generation.
Arm Newsroom

~41 % better performance in image processing workloads.
Arm Newsroom

Significant improvements in super-sampling and speech-to-text tasks.
Arm Newsroom

These improvements make Immortalis-G925 a compelling platform for AI-centric applications such as real-time image enhancement, immersive AR/VR effects, and high-performance AI helpers on smartphones.

Source: Immortalis-G925: A Flagship Mobile GPU (Rockchips.net)