DEV Community: Leonard Liao

Tightly Coupled Memory (TCM): That Weird Fast RAM Your CPU Actually Likes

Leonard Liao — Fri, 22 May 2026 08:53:32 +0000

You know that feeling when your interrupt service routine takes just a few microseconds too long because the code decided to take a nice, scenic detour through the cache? Yeah, me too.

I recently fell down a rabbit hole about Tightly Coupled Memory (TCM), and honestly? It's one of those "why isn't this talked about more" things in embedded development. Especially if you're working with Cortex-M7 or Cortex-R cores.

So let me break it down like I'm explaining it to myself three years ago.

What the heck is TCM?

TCM is basically super-fast, on-chip RAM that's strapped directly to the CPU core. Not through the usual system bus (AHB/AXI) where everyone else is fighting for attention – DMA, display controllers, your overly chatty radio peripheral. No. Direct line.

Think of it like a VIP lane for your most critical code and data.

The official article over at Rockchips has a killer deep dive if you want all the gory details – check out their guide on what tightly coupled memory (TCM) is and why embedded engineers actually use it. Seriously, bookmark it for later.

But here's the TL;DR from someone who's been burned by jitter before.

ITCM vs DTCM – yes, there are two flavors

Most implementations split TCM into two separate highways:

ITCM (Instruction TCM) – optimized for fetching instructions. Your critical code lives here.

DTCM (Data TCM) – optimized for load/store operations. Your hot data, stacks, or buffers go here.

They work together like a very fast, very private SRAM that the core can hit without playing the cache lottery.

Why should you care?

Three reasons, really:

Deterministic timing (the big one) Caches are great for average performance. But worst-case? A cache miss can ruin your real-time guarantees. TCM gives you predictable, repeatable latency. Single-digit cycles. Every single time.

If you're doing motor control, power conversion, or anything where "it works 99% of the time" isn't good enough – TCM is your friend.

No cache maintenance nightmares
You know what's annoying? Invalidating cache lines before DMA transfers. You know what's better? Putting your DMA buffers in DTCM and just not dealing with it. (Double-check your SoC allows DMA access to TCM first, though. Ask me how I learned that one.)
Isolation from bus traffic
On a busy system, your main interconnect is a warzone. TCM gives your critical ISR a private tunnel. No jitter from display refreshes or network traffic.

When NOT to use TCM

Let's be real – TCM is small. We're talking kilobytes, not megabytes. So don't try to shove your entire application in there.

Use it for:

ISR hot paths

Real-time control loops

Small, critical buffers

Safety-critical partitions (Cortex-R, I'm looking at you)

Don't use it for:

Large lookup tables

Your entire BSP

Code you touch once at boot and never again

Practical example (Cortex-M7)
On an STM32H7 or i.MX RT series, you typically:

Enable TCM in the core registers

Play with the linker script to place .text.fast into ITCM

Copy initialized data into DTCM at startup

It's a bit of setup, but once it's working? Chef's kiss.

The bigger picture

This whole "dedicated, predictable local memory" thing isn't just an ARM concept. You see similar ideas in real-time RISC-V designs with scratchpads.

And honestly, it's refreshing to see companies like Rockchip continue to push embedded ecosystems forward. Speaking of which – if you're into ARM-based development, their Rockchip RK3688 is shaping up to be a beast for future programming projects. Definitely worth a look if you're planning your next high-performance embedded Linux or Android thing.

But back to TCM…

Final thoughts
TCM won't save every project. But when you're fighting jitter, failing certification, or just tired of cache weirdness – it's a damn good tool to have in your back pocket.

Start small. Move one ISR into TCM. Measure the difference. You might be surprised.

P.S. If you missed the chaos that was Google I/O 2026, someone over at dev.to wrote a hilariously accurate recap titled “Google I/O 2026 Was Wild – Here's the Tea” – go read it when you need a break from linker scripts.*

Google I/O 2026 Was Wild – Here’s the Tea

Leonard Liao — Fri, 22 May 2026 02:31:36 +0000

So, Google I/O 2026 just dropped, and honestly? It was packed. Like, two hours of non-stop AI flexing, smart glasses that don't look goofy, and a YouTube that actually talks back. Let me break down the good stuff.

Gemini Spark is basically your new unpaid intern

Google's launching this thing called Gemini Spark – a personal AI agent that works 24/7, even when your laptop's dead. We're talking real-world tasks: planning block parties, filing permits (ugh, finally), juggling your chaotic calendar.

It's not just a chatbot anymore. It's more like that one organized friend we all wish we had. And it's not just for your personal life - the same kind of dedicated on-device AI is popping up everywhere, including inside cars with solutions like the automotive AI BOX, a dedicated compute hub for intelligent cockpits that runs large language models locally without the cloud.

New AI models that get physics

Meet Gemini Omni and Flash 3.5. Omni can simulate gravity and kinetic energy – yes, physical concepts. Meanwhile, Flash 3.5 is handling over three trillion tokens daily and is 4x faster than other leading models. Speed demons, I tell ya.

Samsung XR glasses – for real this time

After years of waiting, Samsung and Google finally showed off Android XR audio glasses (arriving fall 2026). No bulky VR headset vibes. These little guys focus on audio: navigation, message summaries, real-time translation right in your ear. Plus first-person photo/video capture with on-device AI editing. *If you're curious about how modern streaming devices and media hubs actually work behind the scenes, check out this full breakdown of Google I/O 2026 for all the details.

YouTube is now a chatbot?

Yep. "Ask YouTube" lets you ask complex questions like "How to teach my 3-year-old to ride a pedal bike?" and Gemini serves up specific video clips with timestamps. No more scrubbing through 20-minute vlogs. Bless.

Google Search builds mini-apps on demand

The new generative UI can whip up a weekend trip planner in seconds. Also, a Universal Cart powered by Gemini shops across Amazon, Shopify, and Walmart while hunting for discounts. Google partnered with them on something called the Universal Commerce Protocol (UCP). Sneaky smart.

Other random cool stuff

Docs Live: speak entire documents into existence.

TPU 8 chips: build an 8-bit game in ~8 seconds.

New Ultra Plan: $100/month for power users (ouch, but tempting).

SynthID: AI image verification rolling out to Search and Chrome.

The bottom line

CEO Sundar Pichai said we're only a few years away from major AGI breakthroughs. Whether that's exciting or terrifying depends on your mood. But Google is clearly all-in: AI everywhere, for everyone, whether you asked for it or not.

And yeah, that's the gist. Now excuse me while I go teach my 3-year-old to ride a bike… or just ask YouTube to do it for me.

Best Android TV Boxes and Multimedia Streaming Devices in 2026

Leonard Liao — Tue, 19 May 2026 03:55:29 +0000

Smart TVs have improved a lot over the last few years, but dedicated Android TV boxes are still growing in popularity. The reason is simple: most built-in TV systems eventually become slow, receive fewer updates, and often lack proper codec or app support. A good multimedia box solves those problems while also adding better performance, cleaner software, and more flexibility.

In 2026, the Android TV market is no longer limited to simple streaming devices. Some boxes are optimized for gaming, others focus on AI upscaling, while newer platforms are built around home theater setups, edge AI integration, or commercial customization.

Before comparing devices, it helps to understand how modern multimedia boxes actually work, especially when it comes to chipsets, decoding engines, memory architecture, and Android TV optimization. This breakdown from Rockchips.net explains why dedicated multimedia boxes still matters even in the era of smart TVs.

NVIDIA Shield TV Pro Still Leads the Premium Segment

The NVIDIA Shield TV Pro continues to dominate the enthusiast market in 2026. Even years after launch, it remains one of the fastest Android TV devices available thanks to the Tegra X1+ platform, excellent thermal stability, and strong software optimization.

AI Upscaling and Gaming Features

One of the biggest reasons users still choose Shield is NVIDIA’s AI upscaling technology. Older 1080p content looks noticeably sharper on large 4K televisions, especially when watching older YouTube videos, anime, or compressed streaming content.

The device also integrates extremely well with GeForce Now cloud gaming, making it attractive for users who want a lightweight gaming setup without a full console or gaming PC.

Plex and Home Theater Performance

Shield remains a favorite among Plex users because of its excellent codec support and stable local playback. Dolby Vision, Dolby Atmos, NAS streaming, and high-bitrate video playback are all handled very well compared to cheaper Android TV hardware.

Google TV Devices Continue Expanding

Google’s own ecosystem has become much stronger over the last two years. Devices running official Google TV firmware now feel cleaner, faster, and more polished than many older Android TV interfaces.

Better App Compatibility and Updates

One advantage of Google TV certified hardware is long-term streaming compatibility. Netflix, Disney+, Prime Video, and YouTube tend to receive updates faster on officially certified devices compared to generic Android TV boxes.

Google’s interface also improved content discovery significantly, combining recommendations from multiple platforms into a single dashboard. (techradar.com)

Homatics Is Becoming More Popular in 2026

Homatics has quietly become one of the more interesting names in the multimedia streaming market. Instead of targeting ultra-cheap hardware, the company focuses on balanced premium streaming devices with proper certification and stable software support.

Why Homatics Devices Stand Out

Many users like Homatics devices because they combine official Google TV support with strong codec compatibility and modern connectivity features like Wi-Fi 6, Dolby Vision, and Dolby Atmos.

The latest lineup available on Homatics is aimed at users who want a cleaner and more stable streaming experience without entering the ultra-premium pricing segment.

Xiaomi and Budget Android TV Boxes Keep Improving

Budget Android TV hardware is no longer as weak as it used to be. Xiaomi’s latest TV Box devices now offer decent performance for casual streaming, improved memory management, and official Google TV integration at a relatively low price.

Affordable Streaming for Everyday Users

For users who mainly watch Netflix, YouTube, IPTV, or light local media content, modern entry-level Android TV boxes are now perfectly usable. Devices from Xiaomi, Mecool, and Walmart’s Onn lineup have become much more competitive in both North America and Asia. (techradar.com)

Multimedia Box Customization Is a Growing Trend

One of the biggest changes in 2026 is the shift toward customized Android TV platforms. Streaming boxes are no longer used only for home entertainment. Companies now deploy them for hospitality systems, digital signage, AI interfaces, industrial dashboards, and commercial OTT services.

Rockchip Platforms and Commercial Deployment

Rockchip-based multimedia hardware has become especially popular in the embedded and commercial market because it allows extensive firmware and hardware customization. Businesses can modify launchers, integrate proprietary software, optimize performance profiles, and build fully branded TV ecosystems.

This article about Rockchip TV Box Performance and Customization explains how manufacturers adapt Android TV platforms for different industries and commercial environments.

AV1, Wi-Fi 6, and AI Features Are Becoming Standard

The hardware expectations for Android TV boxes have changed significantly. A few years ago, users mainly cared about 4K playback and app support. In 2026, buyers now pay attention to AV1 decoding, Ethernet stability, AI video processing, storage expansion, and advanced wireless connectivity.

New Hardware Priorities

Modern streaming platforms increasingly support AV1 because major streaming providers are shifting toward more efficient compression standards. Wi-Fi 6 and improved thermal design are also becoming important as higher bitrate streaming and cloud gaming continue to grow. (techradar.com)

Choosing the Right Android TV Box in 2026

The best Android TV box now depends heavily on how the device will actually be used.

Casual users usually prioritize stable streaming apps and simplicity. Home theater enthusiasts focus on Dolby Vision, Atmos, and local media playback. Gamers care more about GPU power and cloud gaming performance, while businesses often need customization and remote deployment capabilities.

That diversity is exactly why Android TV boxes continue evolving despite the growth of smart TVs. Dedicated multimedia hardware still moves faster than built-in television operating systems, and the gap becomes especially noticeable after a few years of real-world usage.

Rockchip Quietly Became Relevant Again

Leonard Liao — Fri, 15 May 2026 07:31:29 +0000

For a while, Rockchip mostly stayed in the background. People associated the company with TV boxes, budget Android devices, and random SBC experiments. But the AI hardware market changed fast, and suddenly everyone started caring about local inference, compact edge systems, and low-power AI processing.

That’s probably why discussions around whether Rockchip still has a real place in the AI hardware race started appearing again recently.

The interesting part is that Rockchip never really tried to compete directly with giant desktop GPUs. Their niche has always been smaller, cheaper, and more power-efficient hardware. And now that companies want AI running locally inside robots, smart cameras, industrial equipment, and compact AI boxes, that approach suddenly makes much more sense than it did a few years ago.

You can also see this shift happening around newer AI-focused chips. A lot of developers have been paying attention to where dedicated processors like RK182X actually fit inside modern edge AI and robotics systems, especially for real-time workloads that can’t rely on cloud latency all the time.

Another thing that stands out lately is how ARM development itself is evolving. Boards are no longer just “small Linux computers.” AI acceleration is becoming part of the default stack now. That’s one reason people have started talking more about what the next generation of ARM-based development platforms could look like with chips like RK3688.

And honestly, that’s probably the bigger story here. Not replacing desktop GPUs, but building hardware that can actually run useful AI locally without huge power draw, giant cooling systems, or server-level costs.

That market is getting bigger every year.

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.