Hardware Platforms

Roger Cheng

Google Cardboard and Friends

Almost ten years ago, an Oculus Rift DK2 (Development Kit 2) gave me an exciting peek into consumer-grade virtual reality. I was enthusiastic, but the leading edge of VR technology was still very raw and also very expensive. Trying to make this novel technology more accessible, Google Cardboard was a way to turn Android phones into VR headsets. A simple box so cheap, they can be given away as promotions. I have a BB-8 themed viewer that promoted Star Wars: The Force Awakens.

The downside of using a phone is that we only have an accelerometer to sense device orientation. There’s nothing to sense device position. This meant visuals can rotate in response to a head tilt in roll/pitch/yaw directions (three degrees of freedom or 3DOF) but doesn’t change if we take a step left/right, or a step front/back, or sit/stand/kneel. (Which constitute an additional three degrees of freedom for a total of six or 6DOF.)

I eventually decided trading off three degrees of freedom for low cost was false economy. My virtual reality “Ah-ha” moment of leaning in close to a panel was impossible to do in a 3DOF system like Google Cardboard. It’s not just a matter of missing features: I quickly get motion sickness in 3DOF VR. No matter how I tried to keep my body still, there are small movements in the remaining three degrees of freedom and after a few minutes my body started protesting the lack of visual feedback for that motion.

Still, the price was low, which translated to high distribution volume. People tried to iterate on the idea to grow the market, and I kept hoping I could find something I like. Spending money that I should have saved towards a real 6DOF VR system.

The most entertaining take was a VR revival of the View-Master brand. I had an old-school View-Master with a few picture discs, and that nostalgia motivated me to buy one of these new viewers. Technologically speaking it was merely Google Cardboard in View-Master’s signature red plastic, including the orange “lever”. As it was merely a styling and software effort, the business case failed: VR content cost a lot more to produce than those old View-Master picture discs! The best thing I can say is the fact View-Master experiences were only good for short durations, avoiding my motion sickness issue.

With big brands like Mattel and Google onboard, a lot of other brands jumped into the market looking for a successful niche. This was a “Utopia 360” viewer that added two axes of adjustments to improve visual comfort: (1) focal distance between our eyes and the phone, and (2) IPD adjustment. (Interpupillary Distance, or the distance between eyeballs.) Instead of standard tap-on-screen interface, this viewer bundled a small Bluetooth controller. Unfortunately, these features needed software-side support to be useful, and approximately nobody bothered to do so. (This particular unit had a troublesome spring-loaded generic phone holder, so I decided to make a custom holder as one of my first 3D printing projects.)

Samsung is never shy about throwing money at experimental niches. They took a stab with the Gear VR. Going beyond standard Google Cardboard, Samsung added a directional keypad to the side as well as higher quality accelerometer for faster and more accurate 3DOF feedback. I didn’t have a Samsung phone but had a friend who had a Galaxy S7. I thought he shared my enthusiasm of VR, but I later learned he was just being polite while I spewed my enthusiasm. How did I learn this? I bought this Gear VR for him to use with his phone. Years later, he retired that S7 and donated it to my pile of retired Android phones I keep for random projects. Along with the phone he also returned the Gear VR, still unopened in its packaging. By then Samsung has moved on to other things and shut down their Gear VR software support ecosystem so now I can’t do anything with it either.

My final 3DOF VR experiment was this first-generation Google Daydream viewer. It was a small additional expenditure as I already had a Google Pixel phone to go with it. Daydream was Google’s own evolution of the Cardboard concept, with at least two advancements: there were two capacitive touch nubs on the headset to help the phone align its onscreen image. A handheld remote was included, much like the Utopia 360. Google used their muscle to get more software support for Daydream controllers than Utopia 360 ever got for theirs, but there was no way to overcome the fundamental limitations of 3DOF VR.

This string of experiments firmed up my position on virtual reality: 6DOF or GTFO. By the time Oculus released their Go headset, I dismissed it as just another 3DOF system with no meaningful advantages over my Google Daydream. I decided against buying a Go, saving up money towards a 6DOF system of my own.

Roger Cheng

My Virtual Reality “A-Ha” Moment

Nearly ten years ago, I got my first taste of consumer-grade virtual reality hardware when I had the opportunity to put an Oculus Rift DK2 (Development Kit 2) on my head. Up until that point, I had only science fiction stories like Star Trek‘s Holodeck and reading about professional/industrial installations that were priced well beyond my reach. I knew Oculus launched their hardware development as a Kickstarter campaign, but I was too skeptical to put in my own money. I was still very interested in the technology, though, so it would come up in conversation with other tech-oriented friends. I learned one of my friends did pitch in the Kickstarter campaign and was slated to receive a DK2. Unfortunately, my friend’s computer did not meet DK2 GPU hardware requirements and in the absence of data they were reluctant to throw more money at it. I saw an opportunity: my gaming PC had a Radeon HD 7950 GPU which met DK2 minimums. (The minimums would be raised for release, excluding my HD 7950, but that came later.) We decided to meet up and plug their DK2 into my PC so we can both see firsthand what it’s all about.

I have vague memories of software installation struggles mostly with batch files and only a few graphical installer applications. I had to give administrator privileges to many unknown binaries and that made me squirm, and there were error messages to address. All of these unpolished edges were normal and expected of a development kit.

I don’t remember any hardware connectivity issues: I think everything plugged in together just fine. When the picture actually came up, the first impression was rather underwhelming. DK2 display panel resolution was relatively low, resulting in a blurry picture as if my eyeglass prescriptions are out of date. Plus, there was a distracting “screen door effect” caused by visible black lines between pixels. But of course, if we just wanted a static viewpoint, we could have just stared at a computer monitor. Things got more interesting once we started moving our heads to look around, leveraging key elements of virtual reality technology.

The demo applications (all under development) were mixed. It was definitely early days for the technology, with lots of people trying ideas to see what works. There were many standalone test apps and a few VR modes grafted onto existing titles. My friend and I quickly reached agreement we didn’t care for the titles that simulated motion independent of our seating position. The worst of those were roller-coaster simulations, one of them caused my friend to loudly proclaim “NOPE!” and yanked the headset off their head. We both got motion-sick from such experiments and had to take a break.

We were starting to think the whole thing might be a waste of time and money when we fired up Elite: Dangerous and its then-experimental VR mode. After our experience with VR roller-coaster and the like, we were not optimistic about flying around in a spaceship. But hey, we’ve come this far, might as well take a look. I remember it took some effort to get the game to switch from computer monitor over to the DK2 headset. My friend fiddling at the keyboard and the DK2 on my head. “Do you see the cockpit yet?” “Nope” “How about now?” “Still nope” Then it came up. “Hey I see something!”

The ship was still unpowered, so the only movement were of my own head. Even then I could look around at the controls and it felt like I was at the controls of a spaceship. A virtual representation of a reality that’s out of my reach: I could go on real rollercoasters; I couldn’t fly real spaceships. This was all very promising, but there was a problem. Elite Dangerous ship cockpits were designed to be shown on high resolution monitors. Sitting in the middle of the cockpit wearing the low resolution DK2 headset, all control labels were blurry and illegible. I suppose if I were already familiar with the game I could go from memory, but I was not familiar with it and didn’t know how to start up my ship.

My friend and I put our brains together, drawing from our collective computer gaming experiences. Maybe pressing “Z” will zoom in? How about the mouse scroll wheel? PgUp/PgDn? Arrow keys? The answer was none of those, because this was something new. I forgot which of us had the insight to lean closer to the panel, but I leaned closer to the labels and found I could read them. Such a simple thing we would do in the real world without thinking, but somehow it took several minutes for us to think of doing it in the VR world.

That was my VR “A-ha” moment. I no longer remember anything from that day after that moment. Did we manage to get our ship into space? Did we get motion sickness from flying around? It didn’t matter. The mundane act of leaning closer to read labels was the moment it clicked in my mind, and I was hooked on the concept of virtual reality. Sadly, I was too cheap to commit to good VR with 6DOF tracking and wasted a lot of money on cheaper 3DOF headsets like Google Cardboard and friends.

Roger Cheng

Hello ESPAsyncWebServer

I have several interesting sensors in my hardware to-do pile, and I think it’d be interesting to do a series of interactive UIs one for each sensor. My first effort in this vein was a browser UI to interactively play with an AS7341 spectral color sensor, and I learned a lot doing it.

One lesson was that I could perform all sensor data computation and visualization in the browser, so I don’t need much in the way of microcontroller capabilities. The ESP32 I used for AS7341 web UI was overkill. All I needed was an I2C bus and WiFi, so I could easily drop down to an ESP8266. However, because I had used the WebServer from ESP32 Arduino Core library, I will need an ESP8266-friendly replacement.

Sticking with Original Repository

I decided to look into ESPAsyncWebServer, which I had been using indirectly as a dependency of ESPHome. In addition to supporting ESP8266 as well as ESP32, it had a few other features I liked. The most important being its ability to serve files from flash file system, so I don’t have to do silly things like hexadecimal encoding files to program memory. Using this capability also means switching from Arduino IDE to PlatformIO, because the latter has easy tools to work with flash file system. And finally, I was intrigued by the claim ESPAsyncWebServer can serve GZip-compressed files. Making better use of our precious megabyte of flash memory.

The only worry is that the GitHub repository looks stale. There are over 100 open issues, and there are 69 pull requests sitting unmerged. Maybe I should use one of the forks that saw more recent development? Since I was introduced via ESPHome, I thought I would try their fork of the repository. It has seen more recent work, but unfortunately as of this writing, the most recent merge has a C syntax error.

.pio\libdeps\d1_mini\ESPAsyncWebServer-esphome\src\AsyncWebSocket.cpp: In function 'size_t webSocketSendFrame(AsyncClient*, bool, uint8_t, bool, uint8_t*, size_t)':
.pio\libdeps\d1_mini\ESPAsyncWebServer-esphome\src\AsyncWebSocket.cpp:105:23: error: expected primary-expression before ')' token
  105 |   if (!client->send();) return 0;
      |                       ^

It’s been sitting broken for over a month now. I don’t know the story behind the scenes, but it is clear the repository is in no shape to be used. I don’t know which of the other forks might be worth investigating. As an introduction, I’ll start with the original until I run into a compelling reason to do otherwise.

SPIFFS vs. LittleFS

There were a few examples to help me orient myself with ESPAsyncWebServer. Compiling them myself, though, brought up warnings that SPIFFS is now deprecated and I should switch to LittleFS. This was opened as an issue #780 against ESPAsyncWebServer library, but was closed without further action because the warning came from an optional SPIFFS-specific component called SPIFFSEditor. Since it is optional and not relevant to my projects, I can choose to ignore the warning.

Switching over is a little tricky because we build the code and file system separately, and they are two different uploads. If I upload the file system first, it gets reformatted by the code when it sees a file system it doesn’t recognize. As per documentation: “attempting to mount a SPIFFS volume under LittleFS may result in a format operation and definitely will not preserve any files, and vice-versa” In order to end up with a functional system, I need upload code using LittleFS then upload LittleFS file system contents.

Gzip Compressed Files

After verifying a simple “Hello World” web server worked, I compressed the small index.html into index.html.gz. Uploading the new LittleFS system, I was happy to see it come up in my browser! But how does this work? I queried using curl and found the answer:

HTTP/1.1 200 OK
Content-Length: 1284
Content-Type: text/html
Content-Encoding: gzip
Content-Disposition: inline; filename="index.html"
Connection: close
Accept-Ranges: none

ESPAsyncWebServer added “Content-Encoding: gzip” to the HTTP header effectively making the browser deal with it. As the browser will be running on a far more powerful CPU than the ESP8266, it is indeed better suited to handle decompression.

Serve Angular Web Application

As a test of compression, I brought over my Angular tutorial app “Tour of Heroes”. It’s not as simple as just copying the files, though. As per Angular deployment guide, I needed to add a rule so ESPAsyncWebServer will serve up index.html rather than a 404 error when an URL is not found.

  server.onNotFound([](AsyncWebServerRequest *request){
    request->send(LittleFS, "/www/index.html");
  });

Another problem was that LittleFS has a 31-character name limit. (32 characters + null termination.) Unfortunately, my Angular app bundle had a file polyfills.716cc9a230a87077.js.gz which is 32 letters! The hexadecimal hash was generated by “–output-hashing” option of “ng build”. It’s changed on every build in order to avoid using stale cached versions of polyfills.js. I could skip that feature, but then I run the risk of stale cached files. The other workaround is to skip compressing polyfills.js. Dropping the “.gz” extension resulted in a 29-letter filename that works for LittleFS.

With all files in my Angular app compressed (except for polyfills.716cc9a230a87077.js) the size of the app dropped from ~800kB down to ~230kB. This is a dramatic difference when default available flash storage is only 1MB out of a total of 4MB total flash memory onboard.

Should I use Angular to build my ESP8266-hosted sensor web UI projects? No. It is far too heavyweight and a simple sensor UI does not need the capabilities of Angular application framework.

Will I use Angular to build my ESP8266-hosted sensor web UI projects? Possibly. Angular is not the best tool for the job, but it might be valuable for me to practice Angular by starting with something simple.

Source code for this exploratory project is publicly available on GitHub

Roger Cheng

Potential Small PC Explorations

I had fun playing with the GMKtec NucBox3, an interesting and capable little PC more affordable than Intel’s NUC product line, naturally with some expected tradeoffs for its lower cost. I learned about these little PCs from a Newegg advertisement and, between the time I ordered one and its arrival, I had a failed USB external drive that I transplanted into a small form factor Dell PC. Computers in these two projects represent a spectrum that I should keep in mind for future project possibilities. Which one I buy would depend on a project’s requirements.

Intel NUC

A genuine Intel NUC would be more expensive than any of the other options below, but sometimes it’s worth spending that money. For example, if I’m building a solution that needs to be reliable, I will pay more for a brand name. Or if I want to design something that can be repeated by others, it’s easier for someone to buy an identical Intel NUC than to find, say, a GMKtec. For this reason: If my Sawppy rover ever changes over to an x86-64 PC ROS brain, the official recommended hardware will be an Intel NUC. (Supplemented with suggestions on what to look for in lower-cost alternatives like the NucBox3.)

Just Below $90

But when we’re feeling adventurous and not particularly motivated to pay for quality or consistency, we can go bargain hunting. Searching for various options, I observed a price floor somewhere in the $80-$90 range. I see an interesting hint of economic factors at play preventing things from much lower than $90, but I don’t know what they might be. (As a point of comparison, Raspberry Pi 4 8GB MSRP is $75.)

Lowest Bidder du Jour

Amazon categorized these products under: “Electronics” > “Computers & Accessories” > “Computers & Tablets” > “Desktops” > “Minis”. Sorting them by price today, I see several options right around $89, roughly 40% discount from the price of a NucBox3. To get to that price point we have to give up many things. For example, this item (*) made some notable tradeoffs:

  • Memory is half the size (4GB vs. 8GB), uses older technology (DDR3 vs. DDR4), and is soldered in never to be upgraded.
  • Storage is half the size (64GB vs 128GB), uses much slower technology (eMMC vs. SATA) and is also permanently soldered. However, it does have a 2.5″ SATA bay, which the NucBox3 does not.
  • CPU is three years older and from a different product generation (Celeron J3455 vs. J4125) and it doesn’t meet hardware requirements for Windows 11.

On the upside, it still meets all my hard requirements for robot brain: 64-bit CPU running x86-64/amd64 instruction set, gigabit Ethernet port, small, lightweight, and might run on battery power. Depending on future project requirements, I may choose these tradeoffs in favor of a <$90 bargain.

Buying Refurbished

Looking at inexpensive PCs on Amazon, I saw a lot of refurbished units. Clicking around a few listings, I learned Amazon had set up an entire department. “Amazon Renewed” is dedicated to refurbished products of all kinds, not just computers. I should definitely keep this in mind as an option. Given my personal experience, I’d restrict my search to refurbished Dell products from their corporate line. Which would still leave me with very many options. Check out these guys, each offered at a few bucks under $90:

  • Optiplex 3040 Micro Desktop (*) are bigger than an Intel NUC, but tiny compared to anything else. Skimming Dell’s manual, I see a 2.5″ SATA bay inside. I also see what looks like a M.2 slot on a picture of its mainboard, but M.2 isn’t called out in the manual as a storage option. I see a gigabit Ethernet port and it accepts power from a DC barrel jack, so there’s a possibility it can be persuaded to run on battery power.
  • Optiplex 790 USFF Desktop (*) are significantly larger. Packing an optical drive on top of a 2.5″ drive bay and AC power supply. No robot battery power for this machine, but dual 2.5″ drives are possible via an optical drive caddy. This could work for TrueNAS replication target if my storage drive is a high capacity 2.5″ laptop hard drive.
  • Optiplex 3020 SFF Slim Desktop (*) is a successor to the Optiplex 960 I repurposed to a TrueNAS replication target, with at least one 3.5″ drive bay and one optical drive bay. This would be my default choice if I need to build another replication target machine.

What if I want a parallel port for LinuxCNC? Sadly, that’s an uncommon enough request I can’t filter on Amazon. But when it comes to refurbished small Dell PCs, Amazon Renewed is not the only game in town. There are plenty of other vendors like PC Liquidations, who offers filtering by parallel port. Resulting in a list of refurbished Dell Optiplex with parallel port starting at, you guessed it, a few dollars under $90. All good options if I want to dedicate a cheap PC to a task, which usually also requires me to set up automatic software updates.

(*) Disclosure: As an Amazon Associate I earn from qualifying purchases.

Roger Cheng

Good First Impressions of GMKtec NucBox3

It was fun to poke around internal hardware of a tiny PC I wanted to investigate for use as a robot brain. This GMKtec NucBox3 I ordered off Amazon (*) is a more affordable variation on the Intel NUC formula, and its price (significantly lower than Intel’s own NUC) makes me more willing to experiment with it.

Decent PC for Windows 11 Home

But before I take any risk with nonstandard usage, I should verify it worked as advertised. The 128GB SATA SSD came installed with Windows 11 Home edition build 21H2. Upon signing in with my Microsoft account, it started the update process to build 22H2. I assumed the machine came with a license of Windows either embedded in the hardware or otherwise registered. Windows 11 control panel “Activation state” says “Windows is activated with a digital license linked to your Microsoft account” which I found to be ambiguous. I shrugged because I plan to use it as ROS brain and, if so, I’m likely to run Ubuntu instead of Windows. And if I wipe the SATA drive with a fresh installation of Windows 11, it sounds like I can log in with my Microsoft account and retrieve its license.

The more informative aspect of Windows sign-in and registration is letting me get a feel of the machine in its default configuration. All hardware drivers are in place with no question marks in device manager. Normal user interface tasks were responsive and never frustrating, which is better than certain other budget Windows computers I’ve tried. A NucBox3 is a perfectly competent little Windows box for light duty computer use.

One oddity I found with the NucBox3 was the lack of a power-up screen letting me change boot behavior with a keypress. When a PC first powers up, there’s typically a prompt telling me to press F12 to enter a menu to select a boot device, or DEL to enter system setup, etc. Not on a NucBox3, though: we always boot directly into Windows. The only way I found to enter hardware menu was from within Windows: under “Settings”/”System”/”Recovery” we can choose “Advanced startup” to boot into a special Windows menu, where I can select “Advanced Options” and choose “UEFI Firmware Settings”. This is expected to be an infrequent activity most users would never do, so I guess it’s OK for the process to be a convoluted.

UEFI Menu

Once I got into UEFI menu for NucBox3 I was surprised by how many options are listed. Far more than any branded (Dell, etc) computer I’ve seen and even more than hobbyist-focused motherboards.

Some of these options like “Debug Configuration” almost feel like they weren’t supposed to ship in a final product. My hypothesis is that I’m looking at the default full menu of options for a manufacturer using this AMI (American Megatrends, Inc) UEFI firmware. Maybe the manufacturer was expected to trim it down as appropriate for their product, and maybe nobody bothered to do that.

Under the “Chipset” menu we have device configuration for many peripherals absent from this device. They’re marked [Disabled] but the menu option didn’t even need to be here. The final line was also the most surprising: a selection for resistors on I2C buses. On one hand, I’ve never seen a PC’s I2C hardware exposed in any user-visible form before. On the other hand, if I can figure out where SDA/SCL lines are on this motherboard, maybe I can really have some fun. Why bother with a Raspberry Pi or even an ESP32 to bridge I2C hardware if I can attach them directly to this PC?

Ubuntu Server

But all those shiny lights in UEFI menus were just a distraction. What I really want right now is to control boot sequence so I can boot from a USB flash drive to install Ubuntu Server 22.04 LTS. I found I could do it from the “Boot” section of UEFI menu. Ubuntu Server was installed, it worked, and that was no surprise. A computer competent at a full Windows 11 GUI rarely has a problem with text-based network-centric Ubuntu Server, and indeed I had no problems here. For a ROS brain I would want gigabit networking, all four CPU cores, RAM, storage, and USB peripherals. They are all present and accounted for, even sleep mode if I want to put a robot to sleep.

Variable Input Voltage

The next experiment was to see if this computer is tolerant of variable DC supply voltage. On paper it requires 12V DC and the supplied AC adapter was measured at 12.27V DC. I could buy a boost/buck converter that takes a range of input voltages and output a steady 12V (*) but it would be more efficient to run without such conversion if I could get away with it. Since the NucBox3 used a standard 5.5mm OD barrel jack for DC power input, it was easy to wire it up to my bench power supply. I found it was willing to boot up and run from 10VDC to 12.6VDC, the operating voltage range of 3S LiPo battery packs.

Good ROS Brain Candidate

This little computer successfully ran Ubuntu Server on (simulated) battery power. It handily outperforms the Dell 11 3180 I previously bought as ROS brain candidate and is much more compact for easy integration on robot chassis. Bottom line, I have a winner on my hands here!

I’m glad that Newegg advertisement made me aware of an entire ecosystem of inexpensive small PCs. I need to keep this product category in mind as candidates for potential future projects. I have many options to consider, depending on a project’s needs.

(*) Disclosure: As an Amazon Associate I earn from qualifying purchases.

Roger Cheng

Looking Inside GMKtec NucBox3

I thought the GMKtec NucBox3 looked interesting (at least on paper) as candidate ROS brain, so I ordered one (*) for a closer look despite some skepticism. All pictures on that Amazon listing look perfect, I suspected they were all 3D computer renders instead of photos of an actual product. There’s a chance the actual product looked very different from the listing.

The good news: the product is real and for the most part, as depicted in the listing. I find good fit and finish on its plastic enclosure. There is one downside: fingerprints show up very clearly. I had to wipe down the case pretty aggressively for these pictures and I still see greasy smudges. Well, at least you know these aren’t renders! One instance where oily fingerprint smudges are a feature, not a bug.

I see two brass heat-set inserts on the bottom of the case which will be useful for mounting this little box somewhere. They look very small but this is a small lightweight box so it would probably suffice.

Here we also see where actual product differed from product listing rendering. The company website page for NucBox3 showed an access panel to upgrade memory or storage.

But there’s no such access panel on the real thing, and it’s not clear how to get inside without one. Documents in the box consisted of a minimal warranty card in the box and no instruction manual. No matter, the lack of a convenient access panel or a manual shall not deter me from getting inside for a look.

Hiding fasteners under glued-on rubber feet is a common and effective technique. These four fasteners are not symmetrical so, even though the box is a square, we need to remember correct orientation to reinstall.

Without a convenient access door for upgrades, I wasn’t sure what else would differ from listing picture. I was afraid memory and storage would be soldered-in parts, but I was relieved to find they were standard DDR4 RAM and M.2 2280 SSD as advertised. They’re just a tiny bit harder to access without the panel.

Judging by its M.2 keys, we have the option to upgrade this factory-installed SATA M.2 SSD to a higher-performing NVMe M.2 SSD if needed.

What appears to be empty threaded holes (marked with circles) are actually used to secure the CPU heatsink from the other side. (There’s a fourth one under RAM module and not visible in this image.) Four fasteners (marked with squares) secure the motherboard and must be removed to proceed.

The headphone jack protrudes into the enclosure, so we must tilt the mainboard from the opposite side for removal. But we have to be careful because we are limited by length of WiFi antenna wires.

A block of foam keeps WiFi antenna connectors in place, peeling it back allowed the connectors to be released. The antennae themselves appear to be thin sheets glued to the top of the case, similar to what I’ve salvaged from laptops. How securely were they held? I don’t know. I didn’t try to peel them off.

Freed of WiFi wires, I could flip the mainboard over to see a big heatsink surrounded by connectors. As chock-full of connectors as this product already is, I was surprised to see that there are still several provisions for even more connectors on the circuit board. I’m also very fascinated by connectors used here for USB3, HDMI, and DisplayPort. I usually see them oriented flat against the circuit board as typical of laptop mainboards, but without design pressure to be thin, these connectors are standing upright. This is a tradeoff to fit more connectors on the edge of a circuit board, but each connector must go deeper to obtain the necessary mechanical strength to withstand use.

Looking in from the side, the heatsink appears to have a flat bottom. This is good news if I want to mount a different heatsink on this board, possibly with a fan. The flat bottom means I don’t have to worry about sticking out to make thermal contact with other chips or have to cut a hole to clear protrusions. If I want to mount to the same holes, I will have to drill four holes which unfortunately are irregularly spaced but not an insurmountable challenge. All that said, I’m more likely to just point a fan at this heatsink if heat proves to be a problem.

Using this computer as robot brain also means running it on battery power. Nominal power requirements are listed as 12V up to 3A. My voltmeter measured the factory power adapter output at 12.27V. But what can this thing tolerate? I found this chip directly behind the DC power barrel jack, but a search for DC3905 WK1MEG (or WX1MEG) didn’t turn up anything definitive. Texas Instruments has a LP3905 and Analog inherited Linear Technology’s LT3905. Both chips are designed for DC power handling, but neither footprint matches this chip. This might not even be the power management chip, I’m only guessing based on its proximity to the DC barrel jack.

As far as I know, the highest voltage requirement on this PC are USB ports at 5V. On the assumption that nothing on this machine actually needs 12V, then all power conversion are buck converters to lower voltage levels. If true, then this little box should be OK running directly on 3S LiPo power (Three lithium-polymer battery cells in series) which would range from 12.6V fully charged to 11.1V nominal to 10V low power cutoff. I’ll use the power brick that came in the box to verify everything works before testing my battery power hypothesis.

(*) Disclosure: As an Amazon Associate I earn from qualifying purchases.

Roger Cheng

SATA Optical to 2.5″ Drive Adapter

I dusted off an old Dell Optiplex 960 for use as my TrueNAS replication backup target. The compact chassis had a place for my backup storage 8TB 3.5″ HDD extracted from a failed USB enclosure, which is good. But I also need a separate drive for Ubuntu operating system, and that’s where I ran into problems. There was an empty 3.5″ bay and a SATA data socket available on the motherboard, but the metal mounting bracket was missing, and power supply had no more SATA power plugs.

As an alternative plan, I thought I would repurpose the optical drive’s location. Not just its SATA data and power plugs, but I could also repurpose physical mounting bracket with an optical drive shaped caddy for a 2.5″ SATA drive. (*) It wasn’t a perfect fit but that was my own fault for ordering the wrong size.

Examining the caddy after I opened its package, I saw this oddly bent piece of sheet metal. Comparing against the DVD drive, I don’t think it’s supposed to bend like that. I can’t tell if it was damaged at the factory or during shipping, either way metal was thin and easy to bend back into place.

Also comparing against the DVD drive, I realized I bought the wrong size. It didn’t occur to me to check to see if there were multiple different sizes for laptop DVD drives. I bought a 9.5mm thick caddy (*) when I should have bought something thicker possibly this 12.7mm thick unit.(*) Oh well, I have this one in hand now and I’m going to try to make it work.

To install this caddy in an Optiplex 960 chassis, I need to reuse the sheet metal tray currently attached to the DVD drive.

One side fit without problems, but the other side didn’t fit due to mismatched height. This is my own fault.

There’s a mismatch in width as well, I’m not sure this was my fault. I understand the different form factors to be the same width so this part should have lined up. Oh well, at least it is easier to deal with a ~1mm too narrow adapter because one ~1mm too wide wouldn’t fit at all.

There were slots to take the DVD drive’s faceplate. This is for aesthetics so we don’t leave a gaping hole, the eject button wouldn’t work as it is no longer a DVD drive. Unfortunately, faceplate mounting slots didn’t match up, either. This might also be a function of the wrong height, but I’m skeptical. I ended up using the generic faceplate that came with the caddy.

Forcing everything to fit results in a caddy mounted crookedly.

Which resulted in a crooked facade.

Aesthetically speaking this is unfortunate, I should have bought a taller caddy (*) but functionally this unit works fine. The SSD is securely mounted in the caddy, which is now securely mounted to the chassis. And even more importantly, SATA power and data communication worked just fine, allowing me to install Ubuntu Server 22.04 LTS on an old small SSD inside the caddy. And about that old SSD… freeing it up for use turned out to be its own adventure.

(*) Disclosure: As an Amazon Associate I earn from qualifying purchases.

Roger Cheng

Dusting Off Dell Optiplex 960 SFF PC

After two years of use, my USB3 external 8TB backup drive stopped responding as an external disk. I took apart its enclosure and extracted a standard 3.5″ hard disk drive which seems OK in perfunctory testing. In order to continue using it for TrueNAS replication backup, I’ll need another enclosure. I briefly contemplated getting an USB3 SATA enclosure that takes 3.5″ drives (*) but I decided to use an entire computer as its enclosure: I have an old Dell Optiplex 960 SFF (small form factor) PC collecting dust and it would be more useful as my TrueNAS replication backup machine.

Dell’s Optiplex line is aimed at corporate customers, which meant it incorporated many design priorities that weren’t worth my money to buy new. But those designs also tend to live well past their first lives, and I have bought refurbished corporate Dells before. I’ve found them to be sturdy well-engineered machines that, on the secondhand market, is worth a small premium over generic refurbished PCs.

There’s nothing garish with exterior appearance of an Optiplex, just the computer equivalent of professional office attire. This particular machine is designed to be a little space-efficient box. Office space costs money and some companies decide compactness is worth paying for. Building such a compact box required using parts with nonstandard form factors. For a hobbyist like me, not being able to replace components with generic standard parts is a downside. For the corporate IT department with a Dell service contract, the ease of diagnosis and servicing is well worth the tradeoff.

This box is just as happy sitting horizontally as vertically, with rubber feet to handle either orientation.

Before it collected dust on my shelf, this computer collected dust on another maker’s shelf. I asked for it sometime around the time I started playing with LinuxCNC. I saw this computer had a built-in parallel port, so I would not need an expansion card. (Or I can add a card for even more control pins.) The previous owner said “Sure, I’m not doing anything with it, take it if you will do cool things with it.” Unfortunately, my LinuxCNC investigation came to a halt due to pandemic lockdown and I lost access to that space. TrueNAS replication target may not be as cool as my original intention for this box, but at least it’s better than collection dust.

Even though the chassis is small, it has a lot of nice design features. The row of “1 2 3 4” across the front are diagnostics LEDs. They light up in various combinations during initial boot-up so, if the computer fails to boot corporate IT tech support can start diagnosing failure before even opening up the box.

Which is great, because opening up the box might be hindered by a big beefy lock keeping the side release lever from sliding.

And if we get past the lock and open the lid, we trip the chassis intrusion detection switch. I’ve seen provision for chassis intrusion detection in my hobbyist-grade motherboards, but I never bothered to add an actual intrusion switch to any of my machines. Or a lock, for that matter.

Once opened I find everything is designed to be worked on without requiring specific tools. This chassis accommodates two half-height expansion cards: One PCI and one PCI-Express. On my PCs, expansion endplates are held by small Philips-head screws. On this PC, endplates are retained by this mechanism.

A push on the blue button releases a clamp for access to these endplates.

Adjacent to those expansion slots is a black plastic cage for 3.5″ Hard drive.

Two blue metal clips release the cage to flip open, allowing access to the hard drive. This drive was intended to be the only storage device hosting operating system plus all data. I plan to install my extracted 8TB backup storage drive in this space, which needs to be a separate drive from the operating system drive, so I need to find another space for a system drive.

Most of the motherboard is visible after I flipped the HDD cage out of the way. I see three SATA sockets. One for the storage HDD, one for the DVD drive, and an empty one I can use for my system drive. Next to those slots is a stick of DDR2 RAM. (I’m quite certain Corsair-branded RAM is not original Dell equipment.) Before I do anything else with this computer, I will need to replace the CR2032 coin cell timekeeping battery.

A push on the blue-stickered sheet metal button released the DVD drive. Judging by scratches, this DVD drive has been removed and reinstalled many times.

Putting the DVD drive aside, I can see a spare 3.5″ drive bay underneath. This was expected because we could see a 3.5″ blank plate in the front of this machine, possibly originally designed for a floppy disk drive. The good news is that this bay is empty and available, the bad news is that a critical piece of hardware is missing: This chassis is designed to have a sheet metal tray for installing a 3.5″ drive, which is not here.

I can probably hack around the missing bracket with something 3D-printed or even just double-sided tape. But even if I could mount a small SSD in here, there are no spare SATA power connector available for it. This is a problem. I contemplated repurposing the DVD drive’s power and data cables for a SSD and found adapters cables for this purpose. (*) But under related items, I found a product I didn’t even know existed: an optical-to-hard drive adapter (*) that doesn’t just handle the power and data connectors, it is also a mechanical fit into the optical drive’s space!

(*) Disclosure: As an Amazon Associate I earn from qualifying purchases.

Roger Cheng

Window Shopping: GMKtec NucBox3 Mini PC

A Newegg advertisement sent me down a rabbit hole of tiny little desktop PCs with full x86-64 processors. I knew about Intel’s NUC, but I hadn’t realized there was an entire product ecosystem of such small form factor machines built by other manufacturers. The one that originally caught my attention was distributed by several different companies under different names, I haven’t figured out who made it. But that exploration took me to GMKtec which is either their manufacturer, or a distributor with a sizable collection of similar products built by different manufacturers. The product that originally caught my attention is listed as their “NucBox5” (company website listing and Amazon link *) but I actually found their “NucBox3” (company website listing and Amazon link *) to be a more interesting candidate for my Sawppy Rover’s ROS brain. Both products have a Gigabit Ethernet wired networking port that I demand for resistance against RF interference, but beyond that, their respective designs differ wildly:

First the bad news: the NucBox 3 has an older CPU, the Celeron J4125 instead of the Celeron N5105. But comparing them side-by-side, it looks like I’d be giving up less than 10% of peak CPU performance. There is a huge (~50%) drop in GPU performance, but that doesn’t matter to Sawppy because most of the time its brain wouldn’t even have a screen attached.

A longer list of good stuff balances out the slower CPU:

  • RAM on the NucBox 3 is a commodity DDR4 laptop memory module. That can be easily upgraded if needed, unlike the soldered-in memory on the NucBox 5.
  • They both use M.2 SSDs for storage, but the NucBox 3 accommodates popular 2280 form factor instead of a less common 2245 size used by NucBox 5.
  • The SSD advantage was possible because NucBox 3 has a different shape: is wider and deeper than a NucBox 5, but not as tall. Designed for installation on a VESA 100×100 mount, it will be easier to bolt onto a rover chassis.
  • Officially, NucBox 3 is a fan-less passively cooled machine whereas the NucBox 5 has a tiny little cooling fan inside. (Which I expect to be loud, as tiny cooling fans tend to be.) Given that these are both 10W chips, I doubt NucBox 3 has a more effective cooling solution, I think it is more likely that the design just lets the chip heat up and throttle itself to stay within thermal limits. This would restrict its performance in stock form, but it also means it’ll be easy for me to hack up a quiet cooling solution if necessary.
  • NucBox 5 accepts power via USB-C, which is getting easier and easier to work with. I foresaw no problems integrating it with battery power onboard a Sawppy rover. But the NucBox 3 has a generic 5.5mm barrel jack for DC input power, and I think that’ll be even easier.

A NucBox3 costs roughly 80% of a NucBox5 for >90% of the performance, plus all of the designed tradeoff listed above are (I feel) advantages in favor of the NucBox3. I’m sold! I placed an order (*) and look forward to playing with it once it arrives.

(*) Disclosure: As an Amazon Associate I earn from qualifying purchases.

Roger Cheng

Window Shopping: Mystery Mini PC of Many Names

An interesting item came to my attention via Newegg marketing mailing list for discounts: an amazingly tiny Windows PC. My attention was captured by listing picture showing its collection of hardware ports. Knowing the size of an Ethernet port and HDMI port we can infer this is an itty-bitty thing. Newegg’s specific sale item was generically named “Mini PC” with an asking price of $200. I’m not entirely sure the thing is real: all the images look perfect enough I couldn’t tell if they’re 3D renderings or a highly retouched product photos.

I looked at the other listings by the same vendor “JOHNKANG” and saw several other generically named devices ranging from laptops to external monitors. There were no other similar products, so I think JOHNKANG is a distributor and not the manufacturer of this palm-sized wonder. If JOHNKANG is a US distributor for such merchandise, I guessed they probably have an Amazon listing as well. Sure enough, they have it listed on Amazon also at $200(*) at time of writing. Unlike the Newegg listing, the Amazon listing included this exploded-view diagram showing internals and capabilities.

That’s… pretty darned good for $200! With an Intel Celeron N5105 processor, I see a machine roughly equivalent to capabilities of a budget laptop but without the keyboard, screen, or battery. Storage size is serviceable at 256GB and can be swapped out with another M.2 SSD, though in a less common 2242 format which is shorter than the popular 2280. Its 8GB of RAM are soldered and not easily expandable, but 8GB is more than sufficient for this price point.

A few features distinguish this tiny PC from equivalent-priced laptops, starting with its dual HDMI port where laptops only have one. That might be important for certain uses, but I’m more interested in its wired Gigabit Ethernet port and that it runs on USB-C power input. This machine appears to check off all of my requirements for a candidate Sawppy Rover brain. It’s a pretty good candidate for running ROS slotting just below an Intel NUC in capability but compensates for that with a lower price and smaller physical size. Heck, at this size it is starting to compete with Raspberry Pi and might even fit in a Micro Sawppy.

I found no make or model number listed, which is consistent with a distributor that really doesn’t want us to comparison shop against anyone else who might be distributing the same product for less money. If I want hard details, I might have to buy one and look over the hardware for hints as to who built it. Still, searching for “Mini PC” and “MiniPC” with N5105 CPU found this eBay listing of a used unit with Rateyuso branding. Then I found this AliExpress listing with ZX01 as model name. That AliExpress listing is a mess, showing pictures of several other different mini PCs. Not confidence inspiring and definitely turned me off of buying from that vendor. However, the “ZX01” model name was useful because it led me to this page, which linked to a Kickstarter project that has apparently been taken down due to intellectual property dispute.

Performing an image search using the suspiciously perfect picture/render found the GMKtec Nucbox5(*) which appears to be the same product but with “GMKtec” stamped on top. Looking at the Amazon storefront for GMKtec (*) I see many other small form factor PCs without any family resemblance between their industrial designs. My hypothesis is that GMKtec is a distributor as well, but they have built up a collection of products from different manufacturers and that’s why they all look different. I thought this was encouraging. It implies experience and knowledge with the ecosystem of tiny PCs, offering a breadth of products each making a different tradeoff. I looked over their roster and found one more to my taste.

(*) Disclosure: As an Amazon Associate I earn from qualifying purchases.

Roger Cheng

High Power 600W Power Supply (HP1-J600GD-F12S)

Along with the “keyboard is broken” laptop, I was also asked to look into a mid-tower PC that would no longer turn on. I grabbed a power supply I had on hand and plugged it into the motherboard, which happily powered up. Diagnosis: dead power supply. I bought a new power supply for the PC to bring it back to life, now it’s time to take apart the dead power supply to see if I can find anything interesting. Could it be as easy as a popped circuit breaker or a blown fuse?

According to the label, the manufacturer has the impossibly unsearchable name of “High Power”. Fortunately, the model number HP1-J600GD-F12S is specific enough to find a product page on the manufacturer’s site. The exact model string also returned a hit for a power supply under Newegg’s house brand Rosewill, implying the same device was sold under Newegg’s own name. Amusingly, Newegg’s Rosewill product listing included pictures with “High Power” embossed on the side.

If there is a user-replaceable fuse or a user-accessible circuit breaker, they should be adjacent to the power socket and switch. I saw nothing promising at the expected location or anywhere else along the exterior.

Which meant it was time to void the warranty.

Exterior enclosure consisted of two pieces of sheet metal each bent into a U shape and held together with four fasteners. Once pried apart, I had to cut a few more zip-ties holding the cooling fan power wire in place before I could unplug it to get a clear view at the interior. Everything looks clean. In fact, it looked too clean — either this computer hadn’t been used very much before it blew, or it lived in a location with good air filtration to remove dust.

Still on the hunt for a circuit breaker or a fuse, I found the standard boilerplate fuse replacement warning. Usually, this kind of language would be printed immediately adjacent to a user-serviceable fuse. But getting here required breaking the warranty seal and none of the adjacent components looked like a fuse to me.

Disassembly continued until I could see the circuit traces at the bottom of the board. Getting here required some destructive cutting of wires, so there’s no bringing this thing back online. Perhaps someone with better skills could get here nondestructively but I lacked the skill or the motivation to figure things out nicely. I saw no obviously damaged components or traces on this side, either. But more importantly, now I could see that 120V AC line voltage input wire is connected to a single wire. That must lead to the fuse.

Turning the board back over, I see the line voltage input wire (brown) connected to a black wire that led to a cylinder covered in heat-shrink tubing and held in place by black epoxy. The shape of that cylinder is consistent with a fuse. The heat-shrink and epoxy meant this is really not intended to be easily replaced.

Once unsoldered, I could see the electronic schematic symbol for fuse printed on the circuit board. The “F” in its designation “F1” is consistent with “Fuse”, as do the amperage/voltage ratings listed below. This fuse is a few centimeters away from the caution message I noticed earlier, which was farther away than I had expected. My multi-meter showed no continuity across this device, so indeed the fuse has blown. I cut off the heat-shrink hoping to see a burnt filament inside a glass tube, but this fuse didn’t use a glass tube.

I started this teardown wondering if it was “as easy as a popped circuit breaker or a blown fuse”. While it was indeed a blown fuse, a nondestructive replacement would not have been easy. I don’t know why the fuse on this device was designed to be so difficult to access and replace, but I appreciate it is far better to blow a fuse than for a failing power supply to start a fire.

Roger Cheng

Windows PC Keyboard Beeps Instead of Types? Turn Off “Filter Keys”

A common side effect of technical aptitude is the inevitable request “Can you help me with my computer?” Whether this side effect is an upside or downside depends on the people involved. Recently I was asked to help resurrect a computer that had been shelved due to “the keyboard stopped working.”

Before I received the hardware, I was told the computer was an Asus T300L allowing me to do a bit of research beforehand. This is a Windows 8 era touchscreen tablet/laptop convertible along the lines of a Microsoft Surface Pro or the HP Split X2. This added a twist: the T300L keyboard base not only worked while docked, but it could also continue working as a wireless keyboard + touchpad when separated from the screen. This could add a few hardware-related variables for me to investigate.

When I was finally presented with the machine, I watched the owner type their Windows login password using the keyboard. “Wait, I thought you said the keyboard didn’t work?” “Oh, it works fine for the password. It stops working after I log in.”

Ah, the hazard of imprecision of the English language. When I was first told “keyboard doesn’t work” my mind went to loose electrical connections. And when I learned of the wireless keyboard + touchpad base, I added the possibility of wireless settings (device pairing, etc.) I had a hardware-oriented checklist ready and now I can throw it all away. If the keyboard worked for typing in Windows password, the problem is not hardware.

Once the Windows 8 desktop was presented, I could see what “keyboard stopped working” meant: every keypress resulted in an audible beep but no character typed on screen. A web search with these symptoms found this Microsoft forum thread titled “Keyboard Beeps and won’t type” with the (apparently common) answer to check Windows’ Ease of Access center. I made my way to that menu (as the touchscreen worked fine) and found that Filter Keys were turned on.

Filter Keys is a feature that helps users living with motor control challenges that result in shaky hands. This could result in pressing a key multiple times when they only meant to press a key once or jostling adjacent keys during that keypress. Filter Keys slow the computer’s keyboard response, so they only register long and deliberate presses as a single action. Rapid tap and release of a key — which is what usually happens in mainstream typing action — are ignored and only a beep is played. Which is great, if the user knew how to use Filter Keys and intentionally turned it on.

In this case, nobody knows how this feature got turned on for this computer, but apparently it was not intentional. They didn’t recognize the symptoms of Filter Keys being active. Lacking that knowledge, they could only communicate their observation as “the keyboard stopped working.” I guess that description isn’t completely wrong, even if it led me down the wrong path in my initial research. Ah well. Once Filter Keys were turned off, everything is fine again.

Roger Cheng

Mystery Slot in Xbox Series X Packaging

In the video game console market, I am definitely on Team Green of the pie chart going all the way back to the original Xbox. Right now, the Xbox hardware product line is split into two: the expensive Series X with maximum power and the Series S which made design tradeoffs for affordability. Supplies of both were hampered by global electronics supply chain disruption at launch. I wanted a Series X but I didn’t want it badly enough to pay a scalper premium. The Series S got sorted out and has been widely available for several months, and I was happy to find that the supply of Series X is just starting to catch up to demand. During this year’s Black Friday sales season when everyone was out looking for discounts, I was just happy to find Series X available at all for list price. (There were discounts on Series S, but I was not interested.)

When I flipped opened the box, I was happy to see that Microsoft put some design effort into its packaging. The unboxing experience isn’t up to the premium bar set by Apple & others but a far step above the “sufficient and practical” packaging of past Xbox consoles. The console itself is front and center, wrapped like a gift under a “Power Your Dreams” banner. A cardboard box behind the console held a power cable, a 4K120FPS capable HDMI cable, and a single Xbox controller complete with a pair of AA batteries.

Underneath the console, between two blocks of packaging foam, is a piece of cardboard. This turned out to be a “Getting Started” card for those too impatient to read a manual.

The bottom of that card has a fold, and a rounded slot was cut out of it. Why is it shaped like that?

Making this fold and cutting out that slot consumed manufacturing time and money. This was definitely an intentional design choice, but I can’t tell what its purpose could be. The information printed on the card is specific to Series X, so the cutout wouldn’t have been used elsewhere and just went unused here. I thought maybe it was supposed to help hold it somewhere in the box so we could see it when we flipped it open, but this card was just lying in the bottom of my box. The “Power Your Dreams” banner is front and center, so that’s not the location for this card and (2) I don’t see anything for that slot to fit onto elsewhere.

The rest of the package is too well thought-out for this slot cutout to have been an accident, yet it went unused. I can smell a story here, and I am fascinated, but I have to accept that I will never know the answer.

Roger Cheng

MageGee Wireless Keyboard (TS92)

In the interest of improving ergonomics, I’ve been experimenting with different keyboard placements. I have some ideas about attaching keyboard to my chair instead of my desk, and a wireless keyboard would eliminate concerns about routing wires. Especially wires that could get pinched or rolled over when I move my chair. Since this is just a starting point for experimentation, I wanted something I could feel free to modify as ideas may strike. I looked for the cheapest and smallest wireless keyboard and found the MageGee TS92 Wireless Keyboard (Pink). (*)

This is a “60% keyboard” which is a phrase I’ve seen used two different ways. The first refers to physical size of individual keys, if they were smaller than those on a standard keyboard. The second way refers to the overall keyboard with fewer keys than the standard keyboard, but individual keys are still the same size as those on a standard keyboard. This is the latter: elimination of numeric keypad, arrow keys, etc. means this keyboard only has 61 keys, roughly 60% of standard keyboards which typically have 101 keys. But each key is still the normal size.

The lettering on these keys are… sufficient. Edges are blurry and not very crisp, and consistency varies. But the labels are readable so it’s fine. The length of travel on these keys are pretty good, much longer than a typical laptop keyboard, but the tactile feedback is poor. Consistent with cheap membrane keyboards, which of course this is.

Back side of the keyboard shows a nice touch: a slot to store the wireless USB dongle so it doesn’t get lost. There is also an on/off switch and, next to it, a USB Type-C port (not visible in picture, facing away from camera) for charging the onboard battery.

Looks pretty simple and straightforward, let’s open it up to see what’s inside.

I peeled off everything held with adhesives expecting some fasteners to be hidden underneath. I was surprised to find nothing. Is this thing glued together? Or held with clips?

I found my answer when I discovered that this thing had RGB LEDs. I did not intend to buy a light-up keyboard, but I have one now. The illumination showed screws hiding under keys.

There are six Philips-head self-tapping plastic screws hidden under keys distributed around the keyboard.

Once they were removed, keys assembly easily lifted away to expose the membrane underneath.

Underneath the membrane is the light-up subassembly. Looks like a row of LEDs across the top that shines onto a clear plastic sheet acting to diffuse and direct their light.

I count five LEDs, and the bumps molded into clear plastic sheet worked well to direct light where the keys are.

I had expected to see a single data wire consistent with NeoPixel a.k.a. WS2812 style of individually addressable RGB LEDs. But label of SCL and SDA implies this LED strip is controlled via I2C. If it were a larger array I would be interested in digging deeper with a logic analyzer, but a strip of just five LEDs isn’t interesting enough to me so I moved on.

Underneath the LED we see the battery, connected to a power control board (which has both the on/off switch and the Type-C charging port) feeding power to the mainboard.

Single cell lithium-polymer battery with claimed 2000mAh capacity.

The power control board is fascinating, because somebody managed to lay everything out on a single layer. Of course, they’re helped by the fact that this particular Type-C connector doesn’t break out all of the pins. Probably just a simple voltage divider requesting 5V, or maybe not even that! I hope that little chip at U1 labeled B5TE (or 85TE) is a real lithium-ion battery manage system (BMS) because I don’t see any other candidates and I don’t want a fiery battery.

The main board has fewer components but more traces, most of which led to the keyboard membrane. There appears to be two chips under blobs of epoxy, and a PCB antenna similar to others I’ve seen designed to work on 2.4GHz.

With easy disassembly and modular construction, I think it’ll be easy to modify this keyboard if ideas should strike. Or if I decide I don’t need a keyboard after all, that power subsystem would be easy (and useful!) for other projects.

(*) Disclosure: As an Amazon Associate I earn from qualifying purchases.

Roger Cheng

Google AIY Vision Bonnet & Accessories

The key component of a Google AIY Vision kit is the “Vision Bonnet”, a small circuit board to sit atop the Raspberry Pi Zero WH bundled in the kit. In addition to all the data interface available via the standard Raspberry Pi GPIO pins, this peripheral also gets “first dibs” on raw camera data. The camera itself is a standard Raspberry Pi Camera V2.1 but instead of connecting directly to the Raspbery Pi Zero, the camera cable connects directly to the vision bonnet. There is then a second cable connecting from vision bonnet to the Raspberry Pi camera connector, for the bonnet to forward camera data to the Pi after Vision Bonnet is done processing it. This architecture ensures the Vision Bonnet will never be constrained by data interface limitations onboard the Pi. It can get raw camera feed and do its magic before camera data even gets into a Pi.

The vision coprocessor on this Vision Bonnet circuit board is a Movidius Myriad MA2450, launched in 2016 and discontinued in 2020. Based on its application here, I infer the chip can accelerate inference operations for vision-based convolutional neural networks that fit within constraints outlined in the AIY Vision documentation. I don’t know enough about the field of machine vision to judge whether these constraints are typical or if they pose an unreasonable burden. What I do know is that, now that everything has been discontinued, I probably shouldn’t spend much more time studying this hardware.

My interest in commercially available vision coprocessors has since shifted to Luxonis OAK-D and related products. In addition to a camera array (two monochrome cameras for stereoscopic vision and one color camera for object detailed) it is built around Luxonis OAK SoM (System on Module) built around the newer Movidius Myriad MA2485 chip. Luxonis has also provided far more software support and product documentation on their OAK modules than Google ever did for their AIY Vision Bonnet.

I didn’t notice much of interest on the back side of AIY Vision Bonnet. The most prominent chip is marked U2, an Atmel (now Microchip) SAM-D.

The remainder of hardware consists of a large clear button with three LEDs embedded within. (Red, green, and blue.) That button hosts a small circuit board that connects to the vision bonnet via a small ribbon cable. It also hosts connectors for the piezo buzzer and the camera activity (“privacy”) LED. The button module appears identical to the counterpart in AIY Voice kit (right side of picture for comparison) but since voice kit lacked piezo buzzer or LED, it lacked the additional circuit board.

Roger Cheng

Google AIY Vision Kit

A few years ago, I tried out the Google AIY “Voice” and “Vision” kits. They featured very novel hardware but that alone was not enough. Speaking as someone not already well-versed in AI software of the time, there was not enough documentation support to get people like me onboard to do interesting things with that novel hardware. People like me could load the default demo programs, we could make minor modifications to it, but using that hardware for something new required climbing a steep learning curve.

At one point I mounted the box to my Sawppy rover’s instrument mast, indicating my aspirations to use it for rover vision, but I never got much of anywhere.

The software stack also left something to be desired, as it built on top of Raspberry Pi OS but was fragile and easily broken by Raspberry Pi updates. Reviewing my notes, I realized I published my notes on AIY Voice but the information on AIY Vision was still sitting in my “Drafts” section. Oops! Here it is for posterity before I move on.

Google AIY Vision Kit

The product packaging is wonderful. This was from the era of Google building retail products from easily recycled cardboard. All parts were laid out and neatly labeled in a cardboard box.

Google AIY online instructions.jpg

There was no instruction booklet in the box, just a pointer to assembly instructions online. While fairly easy to follow, note that instructions were written for people who already know how to handle bare electronic circuit boards. Handle the circuit boards by the edges and avoid touching components (especially electrical contacts) and such. Complete beginners unaware of basics might ruin their AIY kit.

Google AIY Vision Kit major components

From a hardware architecture perspective, the key is the AIY Vision bonnet that sat on top of a Raspberry Pi Zero WH. (W = WiFi, H = with presoldered header pins.) In addition to connection with all Pi Zero GPIO pins, it also connects to the Pi camera connector for direct access to camera feed. (Normal data path: Camera –> Pi Zero. AIY Vision data path: Camera –> Vision Bonnet –> Pi.) In addition to the camera, there is a piezo buzzer for auditory feedback, a standalone green LED to indicate camera is live (“Privacy LED”), and a big arcade-style button with embedded LEDs.

Once assembled, we could install and run several visual processing models posted online. If we want to go beyond that, there are instructions on how to compile trained TensorFlow models for hardware accelerated inference by the AIY Vision Bonnet. And if those words don’t mean anything (it didn’t to me when I played with the AIY Vision) then we’re up a creek. That was bad back then, and now that a few years have gone by, things have gotten worse.

  1. The official Google AIY system images for Raspberry Pi hasn’t been updated since April 2021. And we can’t just take it and pick up more recent updates, because that breaks bonnet functionality.
  2. The vision bonnet model compiler is only tested to work on Ubuntu 14.04, whose maintenance updates ended in 2019.
  3. Example Python code is in Python 2, whose support ended January 1st, 2020.
  4. Example TensorFlow information are for the now-obsolete TensorFlow 1. TensorFlow 2 was a huge breaking change, and it takes a lot of work — not to mention expertise — to migrate from TF1.x to TF2.

All of these factors together tell me the Google AIY Vision bonnet has been left to the dusty paths of history. My unit has only ever ran the default “Joy Detection” demo, and I expect this AIY Vision Bonnet will never run anything else. Thankfully, the rest of the hardware (Raspberry Pi Zero WH, camera, etc.) should have better prospects of finding another use in the future.

Roger Cheng

Old OCZ SSD Reawakened and Benchmarked

In the interest of adding 3.5″ HDD bays to a tower case, along with cleaning up wiring to power them, I installed a Rosewill quad hard drive cage where a trio of 5.25″ drive bays currently sit open and unused. It mostly fit. To verify that all drive cage cable connections worked with my SATA expansion PCIe card (*) I grabbed four drives from my shelf of standby hardware. When installing them in the drive cage, I realized I made a mistake: one of the drives was an old OCZ Core Series V2 120GB SSD that had stopped responding to its SATA input. I continued installation anyway because I thought it would be interesting to see how the SATA expansion card handled a nonresponsive drive.

Obviously, because today intent was to see an unresponsive drive, Murphy’s Law stepped in and foiled the plan: when I turned on the computer, that old SSD responded just fine. Figures! I don’t know if there was something helpful in the drive cage, or the SATA card, or if something was wrong with the computer that refused to work with this SSD years ago. Whatever the reason, it’s alive now. What can I do with it? Well, I can fire up the Ubuntu disk utility and get some non-exhaustive benchmark numbers.

Average read rate 143.2 MB/s, write 80.3 MB/s, and seek of 0.22 ms. This is far faster than what I observed by using the USB2 interface, so I was wrong earlier about the performance bottleneck. Actual performance is probably lower than this, though. Looking at the red line representing write performance, we can see it started out strong but degraded at around 60% of the way through the test and kept getting worse, probably the onboard cache filling up. If this test ran longer, we might get more and more of the bottom end write performance of 17 MB/s.

How do these numbers compare to some contemporaries? Digging through my pile of hardware, I found a Samsung ST750LM022. This is a spinning-platter laptop hard drive with 750GB capacity.

Average read 85.7 MB/s, write 71.2 MB/s, and seek of 16.77 ms. Looking at that graph, we can clearly see degradation in read and write performance as the test ran. We’d need to run this test for longer before seeing a bottom taper, which may or may not be worse than the OCZ SSD. But even with this short test, we can see the read performance of a SSD does not degrade over time, and that SSD has a much more consistent and far faster seek time.

That was interesting, how about another SSD? I have an 120GB SSD from the famed Intel X25-M series of roughly similar vintage.

Average read 261.2 MB/s, write 106.5 MB/s, seek 0.15 ms. Like the OCZ SSD, performance took a dip right around the 60% mark. But after it did whatever housekeeping it needed to do, performance level resumed at roughly same level as earlier. Unlike the OCZ, we don’t see as much of a degradation after 60%.

I didn’t expect this simple benchmark test to uncover the full picture, as confirmed after seeing this graph. By these numbers, the Intel was around 30% better than the OCZ. But my memory says otherwise. In actual use as a laptop system drive, the Intel was a pleasure and the OCZ was a torture. I’m sure these graphs are missing some important aspects of their relative performance.

Since I had everything set up anyway, I plugged in a SanDisk SSD that had the advantage of a few years of evolution. In practical use, I didn’t notice much of a difference between this newer SanDisk and the old Intel. How do things look on this benchmark tool?

Average read 478.6 MB/s, write 203.4 MB/s, seek 0.05 ms. By these benchmarks, the younger SanDisk absolutely kicked butt of an older Intel with at least double the performance. But that was not borne out by user experience as a laptop drive, it didn’t feel much faster.

Given that the SanDisk benchmarked so much faster than the Intel (but didn’t feel that way in use) and OCZ benchmarked only slightly worse than the Intel (but absolutely felt far worse in use) I think the only conclusion I can draw here is: Ubuntu Disk Utility built-in benchmarking tool does not reflect actual usage. If I really wanted to measure performance details of these drives, I need to find a better disk drive benchmarking tool. Fortunately, today’s objective was not to measure drive performance, it was only to verify all four bays of my Rosewill drive cage were functional. It was a success on that front, and I’ll call it good for today.

(*) Disclosure: As an Amazon Associate I earn from qualifying purchases.

Roger Cheng

Rosewill Hard Disk Drive Cage (RSV-SATA-Cage-34)

Immediately after my TrueNAS CORE server power supply caught fire, I replaced it with a spare power supply I had on hand. This replacement had one annoyance: it had fewer SATA power connectors. As a short-term fix, I dug up some adapters from the older CD-ROM style power connectors to feed my SATA drives, but I wanted a more elegant solution.

The ATX tower case I used for my homebuilt server had another issue: it had only five 3.5″ hard drive bays for my six-drive array. At the moment it isn’t a problem, because the case had two additional 2.5″ laptop sized hard drive mount points and one drive in my six-drive array was a smaller drive salvaged from an external USB drive which fits in one bay. The other 2.5″ bay held the SSD boot drive for my TrueNAS CORE server. I did not want to be constrained to using a laptop drive forever, so I wanted a more elegant solution to this problem as well.

I found my elegant solution for both problems in a Rosewill RSV-SATA-Cage-34 hard drive cage. It fits four 3.5″ drives into the volume of a trio of 5.25″ drive bays, which is present on my ATX tower case and currently unused. This would solve my 3.5″ bay problem quite nicely. It will also solve my power connector problem, as the cage used a pair of CD-ROM style connectors for power. A circuit board inside this cage redistributes that power to four SATA power connectors.

First order of business was to knock out the blank faceplates covering the trio of 5.25″ bays.

Quick test fit exposed a problem: the drive cage is much longer than a CD-ROM drive. For the case to sit at the recommended mounting location for 5.25″ peripherals, drive cage cooling fan would bump up against the ATX motherboard power connector. This leaves very little room for the four SATA data cables and two CD-ROM power connectors to connect. One option is to disconnect and remove the cooling fan to give me more space, but I wanted to maintain cooling airflow, so I proceeded with the fan in place.

Given the cramped quarters, there would be no room to connect wiring once the cage was in place. I pulled the cage out and connected wires while it was outside the case, then slid it back in.

It is a really tight fit in there! Despite my best effort routing cables, I could not slide the drive cage all the way back to its intended position. This was as hard as I was willing to shove, leaving the drive cage several millimeters forward of its intended position.

As a result, the drive cage juts out beyond case facade by a few millimeters. Eh, good enough.

Roger Cheng

Notes on “Make: FPGAs” by David Romano

After skimming through a Maker Media book on CNC routing wood furniture, I wanted to see what I could learn from their FPGAs: Turning Software into Hardware with Eight Fun & Easy DIY Projects (*) by David Romano. I was motivated by the FPGA-based badge of Superconference 2019, which had (I was told) a relatively powerful FPGA at its core. But all my badge work were at the software level, I never picked up enough to make gateware changes. Perhaps this book can help me?

My expectations dropped when I saw it was published in February 2016. The book is very honest that things are evolving quickly in the realm of FPGAs and things would be outdated quickly, but I was surprised at how little of the information in the book could be transferred to other FPGA projects.

In the preface, the author explained they had worked with FPGAs in a professional context since the early days (1980s) of the field. Seeing the technology evolve over the years and drop in price into hobbyist-accessible range, this book was written to share excitement with everyone. This is an admirable goal! But there is a downside to a book written by someone who has been with the technology for so long. They are so familiar with concepts and jargons that it’s difficult to get in the right frame of mind to explain things in a way that novices in the field can understand.

As an example of this problem, we only got up to page 15 before we are hit with this quote: “Behavioral models and bus functional models are used as generators and monitors in the test bench. A behavioral model is HDL code that mimics the operation of a device, like a CPU, but is not gate-level accurate. In other words, it is not synthesizable.” That sound is the <WOOSH> of indecipherable words flying over my head.

The hardware examples used in this book are development boards built around various FPGAs from Xilinx. To use those boards, we have a long list of proprietary software. It starts with Xilinx software for the FPGA itself, followed by tools from each development board vendor to integrate with their hardware. This introduces a long list of headaches, starting from the fact Xilinx’s “ISE WebPack” was already a discontinued product at the time of writing, with known problems working under 64-bit Windows. And things went downhill from there.

For reference, the hardware corresponding to projects in the book are:

The project instructions do not get into very much depth on how to create FPGA gateware. After an overly superficial overview (I think the most valuable thing I learned is that $display is the printf() of Verilog) the book marches into using blocks of code published by other people on OpenCores, and then loading code written by others onto FPGA. I guess it’s fine if everything works. But if anything goes wrong in this process, a reader lacks knowledge to debug the problem.

I think the projects in this book have the most value for someone who already has one of the above pieces of hardware, giving them instructions to run some prebuilt gateware on it. The value decreases for other boards with Xilinx FPGA, and drops to nearly nothing for non-Xilinx FPGA. This book has little relevance to the Lattice ECP5 on board Superconference 2019 badge, so I will have to look elsewhere for my FPGA introduction.

(*) Disclosure: As an Amazon Associate I earn from qualifying purchases.

Roger Cheng

Adafruit SSD1305 Arduino Library on ESP8266

Thanks to Adafruit publishing an Arduino library for interfacing with SSD1305 display driver chip, I proved that it’s possible to control an OLED dot matrix display from a broken FormLabs Form 1+ laser resin 3D printer. But the process wasn’t seamless, I ran into several problems using this library:

  1. Failed to run on ESP32 Arduino Core due to watchdog timer reset.
  2. 4 pixel horizontal offset when set to 128×32 resolution.
  3. Sketch runs only once on Arduino Nano 33 BLE Sense, immediately after uploading.

Since Adafruit published the source code for this library, I thought I’d take a look to see if anything might explain any of these problems. For the first problem of watchdog reset on ESP32, I found a comment block where the author notes potential problems with watchdog timers. It sounds like an ESP8266 is a platform known to work, so I should try that.

  // ESP8266 needs a periodic yield() call to avoid watchdog reset.
  // With the limited size of SSD1305 displays, and the fast bitrate
  // being used (1 MHz or more), I think one yield() immediately before
  // a screen write and one immediately after should cover it.  But if
  // not, if this becomes a problem, yields() might be added in the
  // 32-byte transfer condition below.

While I’m setting up an ESP8266, I could also try to address the horizontal offset. It seems a column offset of four pixels were deliberately added for 32-pixel tall displays, something not done for 64-pixel tall displays.

  if (HEIGHT == 32) {
    page_offset = 4;
    column_offset = 4;
    if (!oled_commandList(init_128x32, sizeof(init_128x32))) {
      return false;
    }
  } else {
    // 128x64 high
    page_offset = 0;
    if (!oled_commandList(init_128x64, sizeof(init_128x64))) {
      return false;
    }
  }

There was no comment to explain why this line of code was here. My best guess is the relevant Adafruit product has internally wired its columns with four pixels of offset, so this code makes a shift to compensate. If I remove this line of code and rebuild, my OLED displays correctly.

As for the final problem of running just once (immediately after upload) on an Arduino Nano 33 BLE Sense, I don’t have any hypothesis. My ESP8266 happily restarted this test sketch whenever I pressed the reset button or power cycled the system. I’m going to chalk it up to a hardware-specific issue with the Arduino Nano 33 BLE Sense board. At the moment I have no knowledge (and probably no equipment and definitely no motivation) for more in-depth debugging of its nRF52840 chip or Arm Mbed OS.

Now I have this OLED working well with an ESP8266, a hardware platform I have on hand, I can confidently describe this display module’s pinout.