←
Embedded Systems4,666 posts
Embedded Systems avatar

Embedded Systems

@embedsys

The embedded systems newsletter. Subscribe here embedsysweekly.com. Weekly selection of embedded software and hardware articles. Hand-made, on Friday

◷ Joined June 2009▧ 234 Photos and videos
Embedded Systems avatar
Embedded Systems@embedsys·9h

A stop bit isn't really a bit, and a DB-9 is really a DE-9. Hackaday traces serial communication from 20 mA teletype current loops to RS-232 (1960), the MAX232 charge pump, 9600 8N1 framing, null modem cables and USB-to-serial adapters. Includes a useful warning: true RS-232 levels on an MCU RX pin can damage the chip. A nice reminder of why the UART refuses to die. hackaday.com/2026/11/18/seri… #embedded #uart #rs232

Embedded Systems avatar
Embedded Systems@embedsys·11h

Why is x86 little-endian? Because a 1970 terminal did its arithmetic one bit at a time. Gustavo Pezzi's history of x86, from the 4004 to the 486, explains the quirks every assembly beginner calls arbitrary: AX, BX, CX, DX, the parity flag, segment:offset addresses. What I like is the angle: none of it is arbitrary. Each choice was made "under pressure, usually to keep old software running on a new chip". The 8008's instruction set wasn't even designed by Intel but by Datapoint engineers Victor Poor and Harry Pyle, for the Datapoint 2200 terminal, and Datapoint then walked away from the chip. Its bit-serial processing explains the little-endian order; its serial line explains the parity flag, still bit 2 of FLAGS fifty years later. The 8086 itself was a stopgap while Intel waited for the iAPX 432. It aimed at source compatibility with the 8080, so A became AL and HL became BX (hence BX as the only memory pointer among the four), and the odd AX, CX, DX, BX encoding order follows the 8080's BC, DE, HL pairs. Even the separate 256-port I/O space comes from the 8080. A great reminder that backward compatibility is an architecture decision with a 50-year tail. Which legacy choice are you carrying in your own firmware today? pikuma.com/blog/history-of-x…

Embedded Systems avatar
Embedded Systems@embedsys·11h

Take Voyager 1 apart, bay by bay, down to a single memory chip. Raj Shekhar built an interactive explorer of the spacecraft launched in 1977 and still talking to Earth from interstellar space. Open one of its ten bays, then step down through modules and chips, or replay the journey from launch. Hardware history you can click through. Nice! Which chip would you open first? rshekhar.com/voyager/

Embedded Systems avatar
Embedded Systems@embedsys·12h

Comparing FPGAs across vendors has always been painful, partly because "logic cells" mean something different at every vendor. The Gate Array's FPGA finder takes this on: 693 FPGAs and SoC FPGAs from 10 vendors, searchable by LUTs, DSP, block RAM, I/O, transceivers, PCIe and hard processors. What I like most: logic is shown as the vendor publishes it, with the LUT type next to it (LUT4, LUT6, ALM or CPE), plus a CoreScore result where one exists, i.e. how many SERV RISC-V cores really fit. A neutral benchmark is a nice antidote to marketing units. An Efinix Ti2000 listed at 1974K LEs, with no published LUT count, doesn't tell you much on its own. It also covers the smaller players (Cologne Chip GateMate, Gowin, Efinix, NanoXplore, Renesas ForgeFPGA) and has filters for hard RISC-V or Arm cores, on-chip ADCs and flash-based parts. There is a JSON API too. Caveats: the data comes from vendor selection guides, so it is only as good as they are, and CoreScore is still missing for most devices. The selection advice on the page is sound, though: count the I/O in the package you can actually route, not the family maximum, and aim for 60–70% utilization. How do you pick an FPGA today: selection guides, a distributor's parametric search, or habit? thegatearray.com/ #fpga

Embedded Systems avatar
Embedded Systems@embedsys·13h

The instructions for playing Voyager’s Golden Record are engraved on the outside of its cover, where dust will slowly wear them away. A historian has now confirmed from the original blueprints that they are also repeated on the inside. Voyager 1 reached one light-day from Earth on November 18. Glen Swanson, former chief historian of NASA’s Johnson Space Center, chose to study “the hardware, the bottle itself” rather than what’s on the record. He is writing a book with Jon Lomberg, the record’s design director, due in August 2027. What I like is the engineering dilemma. The instructions have to be on the outside so whoever finds the probe sees them first, but the outside is exactly what dust and micrometeorites wear down. The answer is the oldest trick in reliable design: redundancy. Pre-launch photos only hinted at it (a bag was in the way in some), and it took the blueprints to settle it. The second finding is a lesson in configuration management. At first only three records were made: one per spacecraft plus a flight spare. About a month after launch, NASA approved 10 extras, counting the spare. Reports since have put the total anywhere from 6 to 14. Swanson traced them to JPL, the UN, the Library of Congress, the Smithsonian and several NASA centers. Nobody is sure how NASA Headquarters got its copy; it may have come from President Carter. Almost five decades later, nobody was sure how many existed. Would your project’s documentation do better? [your experience?] arstechnica.com/space/2026/1…

Embedded Systems avatar
Embedded Systems@embedsys·16h

Phrack #73 is out, and it has more for embedded engineers than you might expect: QNX userland shellcode, abuse of the Linux IIO subsystem, autonomous underwater vehicles, and a confession about a 2003 Linux kernel backdoor attempt. This is the third release from the new staff: 23 main papers and eight Line Noise articles. Beyond the classic exploitation work (use-after-free, a Chromium sandbox escape through Skia, a Docker escape through the vDSO), a few papers land close to our world. If you ship QNX, shellcode research on its userland deserves a careful read. The IIO subsystem is where many sensor and ADC drivers live, so "semantic abuse" of it is worth understanding. And fortify-headers, a standalone header-based FORTIFY_SOURCE=3, is worth a look for anyone hardening C code on Linux. I also like the history section: the SSH1 CRC32 bug, blue boxing, and the 2003 attempt to backdoor the Linux kernel. The intro ends with a note on AI that I find fair: "As the effort required to put ideas into practice falls, creativity and human intuition forged through experience is valued even higher." (Phrack Staff) My advice: skim the table of contents and start with the paper closest to your product. Which one would you read first? phrack.org/issues/73/introdu…

Embedded Systems avatar
Embedded Systems@embedsys·17h

WiFi 8 won't be ratified until 2027, but Qualcomm's Linux driver patches are already on the kernel mailing list. They are already at v2, and the second version fixes issues raised by an early AI-assisted review (Sashiko). The Ath12k series first targets the QCN9575 802.11bn chipset. For now it only lays the foundation and adds preliminary QCN9575 support. The WiFi 8 features themselves are not wired up yet. Two things stand out to me. First, a vendor posting driver code upstream before the standard is final, rather than shipping an out-of-tree blob later. Second, AI review now shapes patches before human maintainers get deep into them. The real test will be the WiFi 8 feature patches. The headline promise is a theoretical 46 Gbps peak plus better behaviour in noisy environments. phoronix.com/news/Qualcomm-A…

Embedded Systems avatar
Embedded Systems@embedsys·18h

Rust's safety in the kernel has a price, and this time it's measured: writes about 19% slower. RBRD is a new patch series for a RAM-backed Linux block driver written in Rust, a counterpart to the existing C driver, brd. The difference is that RBRD guarantees a read and a write never copy the same page at the same time, which avoids undefined behaviour on the Rust side. Phoronix reports the cost: writes are about 19% slower with one thread and 22% slower with eight. Reads are 2 to 10% slower. The developers' answer: "a block device should never return data that was never written." I like their honesty. They say the gap hasn't been profiled yet, so the lock may not explain all of it. They also share a data point: with one lock for the whole device instead of one per page, 8-thread reads were 20 to 28% slower than brd. This is the Rust-in-the-kernel debate in a nutshell. Is it a Rust tax, or a correctness fix with a benchmark attached? For a RAM disk used for testing and simulations, integrity looks like the right call to me. Would you take the trade? phoronix.com/news/RBRD-Rust-… #rustlang

Peter avatar
⟳ Embedded Systems retweeted
Peter@Peteriishak·Nov 17

This is what real engineering looks like 🔥

Post attachment
Embedded Systems avatar
Embedded Systems@embedsys·19h

A compact half-bridge module built around TI's UCC27211A-Q1 (3.7 A source / 4.5 A sink, integrated bootstrap diode) and two IRF3710 MOSFETs, for loads up to 36 V. It comes with a schematic, BOM and layout files. The inputs are independent, so dead time is up to your MCU's PWM peripheral. allaboutcircuits.com/project… #embedded #electronics #automotive

Embedded Systems avatar
Embedded Systems@embedsys·19h

Method chaining in plain C89, with no preprocessor tricks and no leaks. Sounds impossible? Google's iso-c-chaining shows how it's done. Daniel Miller (Google, AI4SQL team) builds chains like ch_new_String(...)->append(...)->split("_")->sub_slice(1, 4) with C89 features only: • a static "self" pointer in each type's file, turned into an intrusive stack so a Hash can be built inside another Hash • a generic struct as the first field of every type, so any object can be cast to it • reference counting where new objects start at refs = 0, so temporaries consumed mid-chain are freed on the spot It goes as far as a JSON serializer, a SQLite wrapper, an HTTP server, and TypedArrays that avoid wrapping every int in function pointers. It's clever, and a great read on what structs, function pointers and static can do. For embedded, though, I'd be cautious: hidden global state behind self(), a heap allocation on every link of the chain, and objects freed implicitly unless you bind them. The author mentions C11 _Thread_local for threads, but with an RTOS and interrupts, that hidden state is exactly what bites. Would you use this style in firmware, or keep it for host tools? opensource.googleblog.com/20… #c

Embedded Systems avatar
Embedded Systems@embedsys·20h

Custom firmware starts with someone documenting the device. For the Zoom R8, there's now a GitHub repo for that: hardware and firmware information for firmware development, gathered by aturley. This kind of groundwork is what makes it possible to bring a commercial device back under your control. It fits the teardown and reverse-engineering thread we follow in the newsletter, from Dyson battery firmware to the Starlink modchip. Is anyone else here hacking on audio gear firmware? github.com/aturley/zoom-r8-h…

Embedded Systems avatar
Embedded Systems@embedsys·20h

Automate the work, not the responsibility: CELUS's CEO on AI as a copilot for electronics design. A vendor's view, but a fair point. Would you let an agent choose the components for your next board? electronicdesign.com/technol…

Embedded Systems avatar
Embedded Systems@embedsys·Nov 17

Sony Mavica owners, this one is for you: floppy3.5 is a battery-powered WiFi NAS that reads 3.5″ floppy disks. bigcrimping published the build on GitHub. Insert a FAT12 1.44 MB disk and the files in its root folder show up on a web page over WiFi, ready to be copied. I like this kind of project that gives a legacy format a path into today's tools with modern electronics around it. Exposing only the root folder looks like a sensible limit for the use case, since those camera disks are flat anyway. Who still has a box of Mavica floppies somewhere? blog.adafruit.com/2026/11/17…

Embedded Systems avatar
Embedded Systems@embedsys·Nov 17

An internet radio is a good ESP32-S3 project because it touches almost everything at once: Wi-Fi streaming, I2S audio, an I2C display, a rotary encoder, a custom PCB and a 3D-printed enclosure. This CodersCafe build documents each layer, code included. The hardware is simple: an ESP32-S3 N16R8 (16 MB flash, 8 MB PSRAM, chosen for audio buffering), a MAX98357 I2S amplifier driving a 3 W, 4 Ω speaker, an SH1106 OLED and a KY-040 encoder. The PCB was drawn in EasyEDA and the case in Fusion 360. What I liked in the firmware: the encoder isn't handled with a naive edge check. It reads GPIO_IN_REG directly and uses a 4x4 quadrature transition table, acting only once the encoder settles on a detent. A clean answer to the classic bouncy-encoder problem. What I'd question: setup() calls esp_task_wdt_deinit(), so the task watchdog is simply switched off, and the station list and Wi-Fi credentials are hard-coded. Fine for a desk radio, but not habits to carry into a product. Thanks to Shebin Jose Jacob and Nekhil for the article. Would you keep the Arduino framework for this kind of audio project, or go straight to ESP-IDF? hackster.io/coderscafe/inter…