Wasn’t expecting the Raspberry Pi 4 Compute module would lead to an ITX carrier board which packs as much I/O as a Jetson carrier board:
The board also offers two SATA ports, one Gigabit Ethernet port, one RJ45 console port, two HDMI ports, a micro SD card slot, two USB 2.0 ports, as well as the usual 40-pin GPIO expansion header, besides the PCIe x16 slot, a PCIe x1 side slot, and M.2 and mPCIe sockets.
1x 16-lane size vertical PCIe connector with x1 lane implementation in the standard Mini-ITX location
1x 1-lane PCIe side port connector to allow to plug a standard x1 PCIe add-on card on a left side
4x mini PCIe slots with PCIe, USB 2.0, and I2C interfaces
4x M.2 Key E slots with x2 (dual) PCIe lanes, USB 2.0 interfaces
1x M.2 Key M slot to allow M.2 2280 NVME SSD implementing x1 PCIe lane
What I don’t like about this is that, if you are going to make a mini-ITX or any ATX form factor, why not add ATX power input, or at the very least the new ATX 12v standard? You still have to add an external power brick or POE or the other weird connector.
Debian Bullseye
Been running a few tests between Rock64 and Pi as PiOS was moved up to Bullseye. A few interesting differences I found Rock64 tends to stay within 50-60C without thermal throttling, in comparison a Pi without a heatsink on the RAM starts to thermal throttle at 60C. For the most part I was running a few heavy CPU rendering of Mandelbrot, the Spudnik wearable project wouldn’t be crunching that hard however its important to get an idea of thermal factors.
Including a photo of the Rock64 stats, it was weird on rendering tasks that it would put more load on the 2nd and 3rd cores. It does make me wonder a bit how much of a difference it would be if I put a heatsink to the RAM.
Not All Raspberry Pi Hats Work On Other Board(s) GPIO:
There was a hat I had been hoping would work on Pi-compatible boards, it turned out some hats use interrupts differently if they combine multiple functions. Quoting a Pine forum post, “On Pine64 only the PB, PH, PL (and PG) series of GPIOs support interrupts”
Source URL: Mounting RPI3 "hats" on Pine64
Having that info known, it raises the doubts of using a Quartz64 Model B if/when it launches. On the wearable side of things, since the Pi has a camera interface it does trim the power requirements of using a USB 2.0 port. So while Potato/Spudnik Rock64 wearable has GPIO hell to deal with, shifting to a Pi 4 8GB raises the power usage by a small amount–processing wise Pi would make it possible for multi-input via touchscreen and gesture input via sensor array.
3G Shutdown in 2022:
Whenever anybody asks why I don’t like internet enabled vehicles/infotainment/TV/kitchen appliance systems, the 3G shutdown in 2022 is plenty of reasons why. There are also vehicle makers who make “Software as a Service”, if you don’t upgrade your GPS maps pairing a newer phone is impossible. Somehow I hope the chip shortage auto makers are dealing with will lead to a less connected vehicle and maybe modular. From personal experience a Thinkpad of mine had a 3G modem, unlikely a modified BIOS would allow a 4G card.
There are Google services that are almost forgotten they exist, found it interesting Ars Technica did a write up and the Google Toolbar mysteriously shutdown. Last time I remember Google Toolbar being bundled with anything was 2012. Reminds me when I was a web developer in the early 2000s, lots of overlapping APIs by search engines which led to Altavista, Excite & Google being the last big search engines–Altavista ended when Yahoo inked the Microsoft Bing search as their default search engine.
On a related note, in the late 90s Digital Altavista was ahead of its time for having a LAN based search engine which IT could install on a Windows, Linux or Unix server.
Placing a link from Mouser to a TI document on designing a Qi compliant receiver coil. If you’re building/working with embedded hardware that is power sipping its quite possible to attempt Qi based wireless power delivery. Its one thing to experiment in creating a receiver coil, more of a challenge of testing some ideas of how to keep it as compact as possible.
Mouser Link to “TI Designing A Qi Compliant Receiver Coil”: https://www.mouser.com/pdfDocs/TI-Designing-a-Qi-compliant-receiver-coil.pdf
For the past two months I was wondering why lithium battery prices were increasing, there is a raw material shortage going on.
Wearable Computer 2.0 Project
Sourcing a touchscreen hasn’t been easy, just when I think I found a display the flex cable to HDMI is out of stock. I’ve left the option of going without a touchscreen, keeping it less complicated would avoid digitizer driver dependence.
Some changes of making it more modular had been done, the slight adjustment in cable routing makes it possible to swap between a Pine Rock64 and a Raspberry Pi 4. One potential idea is build in the option of an HDMI switch, for switching between a Rock64 running Debian and Ubuntu to get the best of both worlds.
I remember reading about Premise when it first launched about the possibility of users unknowingly doing risky data gathering for foreign gov’ts, quite surprised there hasn’t been that many people detained/arrested for being told to photograph locations. Beyond US/allied countries using Premise, it wouldn’t be shocking if there are other Premise-like apps out there.
URL to the article on Premise:
** NVIDIA launches Jetson AGX Orin Developer Kit**
Jetson AGX Orin has some impressive specs, it’ll be interesting how supply and demand will impact pricing as existing Jetson hardware has gotten just as expensive as GPUs. In terms of MSRP, the first Jetson AGX Xavier 16GB RAM Rev1 was originally $1499 so the increased price of the Jetson AGX Orin to $1999 isn’t that surprising.
CPU – 12-core Arm Cortex-A78AE v8.2 64-bit processor with 3MB L2 + 6MB L3 cache
GPU / AI accelerators
NVIDIA Ampere architecture with 2048 NVIDIA CUDA cores and 64 Tensor Cores @ 1.3 GHz
Storage Benchmarks for Embedded Usage
I’ll be adding benchmarks for storage devices, it might interest others who are running embedded hardware and looking for the best performance on a budget. Much of the projects I’ve been doing on Jetson is navigation & LIDAR. Raspberry Pi has been more of an experimental platform for testing sensors. Rock64 is more or less the middle ground of seeing if a Pi specific project could work on a Pi-like platform.
SanDisk Ultra 16GB MicroSD Class 10 that was included with a Pine64 Rock64(it was originally running a security appliance OS).
Average Read Rate: 47.5 MB/s
Average Write Rate: 13.4 MB/s
Average Access Time: 0.81 msec
Notes: Wasn’t surprising a security company would use the cheapest MicroSD, considering the Rock64 MicroSD slot has a max read speed of ~50-60 MB/s on earlier revision boards a light weight OS is a must. I used this card when I was running a few tests on Armbian Bullseye.
G.Skill 32GB MicroSD Class 10/UHS-1:
Average Read Rate: 80.5 MB/s
Average Write Rate: 5.0 MB/s
Average Access Time: 1.82 msec
Notes: My Kindle Fire rejected this card with an error of it being slow and the benchmarks proved it.
Samsung Evo Select(Green & White version) 64 GB Class 10/UHS-1:
Average Read Rate: 60.5 MB/s
Average Write Rate: 30.2 MB/s
Average Access Time: 1.02 msec
Notes: Awhile back I got tricked by the non-updated specs/product data, originally the 64GB version was UHS-3 while the 32GB version was UHS-1 and Samsung downgraded the 64GB to UHS-1 around the time of the 256GB card came out. Even in a downgraded specs it does perform better for Armbian Bullseye.
SK Hynix 128GB NVMe SSD:
Average Read Rate: 452.2 MB/s
Average Write Rate: 208.1 MB/s
Average Access Time: 0.15 msec
Notes: This is an OEM SSD, even though the write speed is “meh” I’d still consider it usable for data logging in an AI/ML usage task.
SSD Benchmarks PNY 120GB CS900 SATA SSD:
Average Read Rate: 450.9 MB/s
Average Write Rate: 452.0 MB/s
Average Access Time: 0.10 msec
Notes: Benchmark was done with a 100MB test file, with a 500MB test file it benchmarks 451.5 MB/s(read) & 452.5 MB/s(write). While this SSD series is cache-less, it performs reasonably well at the price point. Pre-pandemic this 120GB SSD was on sale for $17.99.
PNY 240GB CS900 SATA SSD:
Average Read Rate: 535.2 MB/s
Average Write Rate: 500.0 MB/s
Average Access Time: 0.10 msec
Notes: This SSD is slightly faster than similar priced models of other makers, when its on sale or has a coupon you’ll find it a solid value. I would note if you read up on reviews the 480+ GB versions the read speeds are closer to the “up to” 560MB/s and writes near 540MB/s. From usage with a Jetson Nano the potential write speed is a bit overkill.
While researching Google Coral & Intel Compute Stick 2 performance, I found the Khadas Vim3’s NPU “Supports a maximum frequency of 800MHz at 5.0 TOPS INT8 inference up to 1536 MAC Internal L2 cache (512KB) and system workspace buffer (1MB) Supports all major deep learning frameworks including TensorFlow and Caffe”. Google Coral’s TPU in comparison is 4.0 TOPS INT8, Intel Compute Stick 2 has the same TOPS as Google Coral but supports more OSes.
There was news last year that Khadas had announced the Vim4 with up to 8GB of RAM, however no exact reference about the NPU even though the Amlogic A311D2 is just an upgraded A311D. The ability to use an NVMe SSD makes the Khadas seem more handy in comparison to some boards out there.
I was only researching the Khadas mostly due to the NPU, from a hardware standpoint having yet another board vs accelerator the latter streamlines workflow. (Pine’s Rock64/Pro leaves a bit to desire at times, Rockchip has better Linux support and is easier to workaround certain things such as enabling Pi compatibility GPIO–Amlogic is more popular for Android/media server boxes than compute)
Khadas Vim4 news article link:
ASRock 4x4 Box-5000 series mini PC with AMD Zen 3 Ryzen 5000 U-Series covers plenty of price points with a Ryzen 3 to Ryzen 7, the 2.5G Ethernet port is going to be a selling point for some. Personally I’d be interested in one of these mini PCs if the Linux support is solid, its also weird these have a better port layout on the front than the ASRock Desk Mini.
Product news article link:
Sky Watch Project
Originally this had been an on & off project, someone I knew had seen a UFO dive at a fast rate of speed into the ocean without any sonic boom off the east coast of the US and used a Raspberry Pi 2 to watch for another possible sighting(s). From a budget side this was more of reusing surplus hardware, had gotten interesting photo/video clips yet the downside was the camera tracking left too much to desire in terms of image quality. Hardware side of things since this remains mostly a side project with minimal work time set aside to this, it’ll likely remain in the focus upon being “keep it simple and stupid”. As far as hardware allocation if it wasn’t for the chip shortage there would of been two Raspberry Pi 3B/3B+ to replace the aging Pi 2 platform.
Adding a Jetson Nano into the mix of hardware happened due to the shortage of chips for the car project, for now it’ll be batch processing photo/video to streamline some of the workflow–the original batch processing was a friends’ Adapteva Parallella. At one point there was consideration of a Parallella cluster however the concept of keep it simple and stupid squashed that move.
Object Classification Experiments
In the past it was about trying to find enough UFO shapes to form a database, YouTube has been interesting for UFO video footage and allowed to build data points of estimating shape/size/speed(s). As far as crunching data a common trend of “Tic Tac” objects is three versions of such craft, a common Tic Tac Type 1 seems to output heat from the aft section, Tic Tac Type 2 seems to shape shift when it speeds up and Tic Tac Type 3 seems hardened for diving into the ocean or flying out of the ocean. Speed wise all three types have speeds that a human wouldn’t be able to handle-that is one area worth noting.
A general conclusion is if you can rule out Lockheed Martin from hypersonic craft that dives into the ocean then it leaves you with Russia as they’ve had a project of a hypersonic aircraft since the Soviet Union era with three potential variants-which was called Project Ayaks(English name: Ajax)
Another possibility is Russia built a craft with a mix of Ayaks and their Yak 141/201, if you mix hypersonic speeds with the VTOL function of a Yak 43, 141 & 201 you’d have something which moves like a Tic Tac UFO. Keep in mind with hypersonic speed you really wouldn’t need a stealth coating as it would be too fast to intercept. Including links below for the Yak 43, 141 & 201. Keep in mind Lockheed Martin did buy VTOL engine performance data of the Yak 43/141 which had some impact in the development of the F-35B. If you’re now wondering why I didn’t include China, well they can’t make reliable engines for their Russian designed planes and their reversed engineering of engines lead to their J-15(Su-33 based) having to fly with less of a payload from their aircraft carriers.
Yak: 43
Sky Watch Moved To Debian Bullseye
Mostly been holding off on posting it as while it was a smooth migration on Pi hardware in general, the original Pi 2 is starting to show its age and some options are replace now with Pi spares then deal with hardware needs later. In terms of Raspberry Pi failure rates its been one Pi 2 and a Pi 3 so pulling a spare from one state for another isn’t a big deal. From a hardware mash up I have my doubts anyone moved to a Pi4 as nobody expected the chip shortage–as for myself I still kept a Pi3B going even though it’ll be shifted to a Pi4.
Project wise I’m considering leaving at the end of the year, it’ll be a conflict of interest to stick around when the Jetson Nano considerations are discussed. From a technical angle Pi 4 A-72 has plenty of power to do everything including Tensorflow Lite, mostly trying to dodge being accused of pushing CUDA and staying hardware neutral.
Rock64 Hardware Migration Plans
Been holding off on openly saying it as I’m waiting on more stable OS builds, however the QUARTZ64 Model-B is more than likely the best hardware for the power profile/usage environment and stick to the 5V 3A power requirements. Model-A seemed overkill for mobile usage even though it had plenty of expansion options and a full sized PCIe slot would go unused.
Compute Project 2.0 ARM Edition
Ever since the chip shortage and crypto hell had created a GPU shortage, I’ve modified some projects to run on ARM. This lead to “Compute Project 2.0” as a potential option of running compute and also simulations on a mix of ARM processors–I’ve been tempted to buy an 8-core ARM server as a test platform but so far using two Rock64 boards showed some interesting promise in an alternative hardware platform. Performance is nowhere near as instant as x64 with CUDA but serviceable for non-overly demanding tasks… for example you could run a compute process to estimate how fast a piece of metal can spin before it shatters for Tesla turbine designing.
Other notes:
This SSD was in a USB external enclosure which supports SSDs. The benchmark on Linux was with the disk utility with a 100MB test file which runs 100 read & write tests.
Speed wise I found the 240GB Kingston A400 to be faster, it really depends upon if you want the extra few GB of storage space and save a few dollars with Microcenter store branded SSDs.
Cement Kayak Project
The concept is the same as the canoe, the only difference is engineering a well designed or purpose built kayak.
Here is a link to Popular Mechanics “How To Make A Concrete Canoe” as a point of reference:
As far as what I’ve been engineering as a side project for the last two years is more or less build it like a two person kayak so it would have similar qualities of a canoe but the aft seat would be for SDR Radio Telescope gear, you could call it a science kayak with plenty of battery power for three days & up to 3-4 days food rations. Due to the chip shortage hell, trying to trim down hardware size/weight is a decision of convenience and also trying to design a modular platform that makes it easy to swap out batteries too. Weight wise since the average person is ~160 lb, a dedicated kayak seat for science would provide plenty of room to work with.
Concept Modules: Lightweight
–Scale the hardware requirements down to a Jetson Nano with large storage(external SSD), plenty enough performance with the benefit of AI/ML functionality for a radio telescope. This lightweight setup would provide the maximum battery runtime for a multi-day kayak travel. High-Performance
–Designed for for clustered compute for radio telescope AI/ML function, originally I would of opted for dual Jetson Xavier NX but with the chip shortage this option is off the table as the NX is impossible to find–possible the pending release of Orin NX is the reason. Another option is up the battery configuration to run a Jetson AGX Orin. (It could be possible to rework CUDA to OpenCL and use a Ryzen Mini PC IGP but it would be less than optimal)
Other notes:
Extra space could support other science tasks.
Turing Pi 2 Cluster Computer
Wasn’t paying much attention to Kickstarter, looks like an ITX/2U rack mount friendly platform that supports Raspberry Pi 4, Turing RK1 and Jetson modules is being made. For raw compute on a budget the RK1 looks like a perfect fit for this cluster board solution https://www.kickstarter.com/projects/turingpi/turing-pi-cluster-board
Article on Hack A Day of testing on a pre-production board:
Storage Benchmark Plans
In the future I’m planning to expand storage benchmarks on NVMe for embedded boards and also portable SSDs since they’re a more popular option for booting a Raspberry Pi 4 these days. As far as anyone wondering why on previous benchmarks I had used cacheless SSDs, in an embedded environment you won’t have the amount of disk activity as with a desktop and easier to meet the typical “Pi Project Budget”.
You haven’t been following Jeff Geerling on youtube? The TP2 is pretty old news by now.
What I don’t like about it is that everything has access to something else. 1 Pi CM4 does display, another does M.2, another does SATA and so on. Unless you do a network boot, you need CM4s with eMMC.
I feel like this board works more like a monolith with multiple non-unified compute devices. If you buy 3 TP2s, or 2 + one passive witness / tie breaker (can be a Pi 2 for all it cares), you may be able to do a cluster, but by that point, you’d probably be better served by other single board computers in a cluster, that can all work independently.
Haven’t been paying any attention to Tech YouTube for awhile mostly due to the Gamers Nexus prebuilt PC tear down videos that were in my feed. Also been busy between PowerPC application layer on ARM and deploying SSDs on an ongoing compute project.
I agree there are better cluster options for working independently, that is what I prefer. Usage wise the way that some want to use a Pi 4 such as an ultra low power file server, it might be useful to have an independent function in certain use cases. If I were trying to setup a low power yet fast file server, I’d pass on using a Pi due to the chip shortage and drop in a 1TB M2 SSD & 2TB SATA SSD into a Mini PC.
For storage, I bought about 2 months ago an Odroid HC4. I haven’t set it up yet to this day, because I’m still learning how to get an OS that is not the official ubuntu, or armbian, or manjaro , properly running on it.
And I’ve had a hard time trying to install ZFS on it, which is why the long delay. But I think it will make for an interesting storage for the rest of my SBCs, I’m planning to netboot most of them.
Using A Chromebook For Work
Many years ago I had gotten a Chromebook which had a built-in 3G modem, originally had bought it to make use of the free 200 MB monthly data plan while out and about if I didn’t have data coverage with my phone. When I read about using Crouton to run Linux applications my interest in offline application usage grew, I learned the Haswell Celeron could squeeze nearly 10 hrs battery run time while offline vs a normal Windows laptop. I later moved onto a quad-core Celeron, battery life with Linux applications boosted run time to 11 hrs and started using a Chromebook more often as the longer battery life was hard to give up.
Usage case of a Chromebook while offline using Linux applications:
-Distraction free work environment
-Battery life on ChromeOS crushes Windows with similar priced/spec-ed systems, the amount of background crap going on within Windows even on a clean install chews battery life. When I was using my Windows based i3 8130U laptop the average battery run time I had was ~6.5 hrs, when I looked at Reddit of Chromebook owners with a similar i3 8130U they were getting close to 8 hrs.
-Build quality of Chromebooks are above average, however the quad-core Celeron I own is considered an education model designed for more wear & tear.
-Older models such as a Celeron N3060 and earlier you can remove the write protect screw and convert the EOL Chromebook to a pure Linux laptop.
Keeping A Project Going With Hardware Delays
Everybody at some point had gotten lazy of relying upon a Raspberry Pi for all their projects. When the chip shortage got worse everybody has jumped to other similar hardware such as Orange Pi which has been fairly reliable with the GPIO compatibility from what I’ve read.
My personal opinion is never ever get yourself into a hardware vendor lock-in, don’t base a project to be processor dependent as mobility to switch hardware gives you plenty of options–Apple is a great example of this, they actively maintained an x86 version of MacOS during the dark period of uncertainty of early PowerPC in the 90s, when Apple bought NeXT they were smart to maintain an x86/x64 development of Darwin/OS X so when Motorola/Freescale hit the wall with the G5 the switch to Intel CPUs was flawless.
Why am I making this point?
I’ve hit the power delivery USB-C issue on the Raspberry Pi 4 quite a few times, it crapped bricks while running on a Power Bank and another time I had issues while running on solar power. When a stupid engineering design flaw gets in your way, find your best options and move on when necessary. In my situation one project I’m jumping ship from had stupidly gotten themselves locked into the Pi due to the lower cost, I can only hope common sense leads to change.
In my own projects while Pine64 Rock64 had been more of a “test” usage case, their community came up with workarounds on the GPIO compatibility configuration files which makes it possible to use a fair number of common/popular Pi hats. I’ve considered moving a project to run on x64, the only reason I haven’t is the lower power side such as a Celeron N4xxx using two USB devices would hit the power delivery limit out of the four ports so the only option is scale the power usage upwards and that would make using an AMD Ryzen IGP’s OpenCL more ideal.
