The BeagleBoard.org Foundation introduced BeagleBone-AI SBC at Embedded World 2019 last February. The board is specifically designed for artificial intelligence workloads at the edge thanks to Texas Instruments AM5729 dual-core Cortex-A15 processor that embeds a dual-core C66x DSP, and 4 EVE (Embedded Vision Engine) cores.
The BeagleBone Black compatible board was not available at the time, but the Foundation has now formally launched the board, and you can buy BeagleBone-AI for $118 and up with heatsink and antenna on sites such as Mouser, OKdo, or Newark.
BeagleBone-AI full specifications have now been published:
- SoC – TI Sitara AM5729 with
- Dual-core Cortex-A15 processor @ 1.5 GHz
- 2x dual-core PRUs
- 2x Cortex-M4 real-time cores
- dual core C66x VLIW DSP
- 4x EVEs
- 2.5MB of on-chip L3 RAM
- VA-HD subsystem with support for 4K at 15fps H.264 encode/decode and other codecs at 1080p60
- Vivante GC320 2D graphics accelerator
- Dual-Core PowerVR SGX544 3D GPU
- System Memory – 1GB RAM
- Storage – 16GB on-board eMMC flash with high-speed interface, MicroSD card slot
- Video & Audio Output – Micro HDMI port
- Connectivity – Gigabit Ethernet and high-speed WiFi, dual-band WiFi 5 and Bluetooth 4.2 LE
- USB – 1x USB 3.0 type-C for power and superspeed dual-role controller, 1x USB 2.0 type-A host
- Expansion – 2x 46-pin BeagleBone Black (BBB) compatible headers working with existing cape add-on boards
- Debugging – Serial header
- Misc – Reset button, 5x user LEDs, power button
- Power Supply – 5V/3A via USB-C port
- Dimensions – 86.4 x 53.4 mm (compatible with BBB)
The solutions is said to offer a “zero-download out-of-box software experience” with TI C66x digital signal processor (DSP) cores and embedded-vision-engine (EVE) cores supported through an optimized TIDL (Texas Instruments Deep Learning) machine learning OpenCL API with pre-installed tools. The Linux powered board targets automation in industrial, commercial and home applications. You’ll find complete documentation (Wiki tab), and hardware design files since the board is open-source hardware, in Github.
Further information may also be found on the product page.
Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.
Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress
I was hoping for such an AM5xxx based BeagleBone 5 years ago. A dual A15 is a bit outdated in 2019 in terms of efficiency, even with embedded AI cores. And the price may make users think twice. It’s true that BB has provides clean products, excellent support and doc, but on high volumes cheaper alternatives may become attractive.
I’m not familiar with these kinds of chips but why anyone wants *2* PRUs when it already got *2* Cortex M4 units? Don’t they serve the same hard real-time purpose?
If I remember driving ws2811 LEDs with the PRU correctly you can do stuff like shift out directly to a GPIO instead of shifting, masking and then updating the GPIO register and other things that might be handy for some applications. It’s also possible they are just there for users that were using the PRUs in another TI chip.
do the m4s have access to shared RAM?
The PRUs are very simple but very fast (200MHz, all instructions are 1 clock cycle IIRC), originally meant for handling industrial Ethernet protocols such as EtherCAT, but in the maker community have been used for tasks such as pulse generation for stepper drivers.
The BeagleBones aren’t a good choice for a TV box, but have unusual features that are a good fit for many industrial applications.
Agreed, I used mine as a JTAG adapter with OpenOCD, it took me one hour to hack a GPIO driver, and it was way faster and more reliable than any parallel port or usb crap I tried in the past!
The solutions is said to offer a “zero-download out-of-box software experience – iam not familiar with beagleboards…what kind of system does it have preinstalled????
If it’s anything like the Beaglebone Black:
– Debian, with a bunch of robotics/computer vision software packages already installed
– Configured to work in USB gadget mode: when plugged into a computer, it enumerates as a serial port, a network adapter and a Mass Storage Device (just shows up as a disk drive, with drivers for windows, and a link to a web page hosted over the network adapter).
It’s a pretty decent setup for getting started.
At $118USD and with an outdated A15 32-bit core, i think this generation of beaglebone will not be as successful as the previous ones, at least in the SBC enthusiast market. For embedded linux developers however, this still is an attractive board; especially with the PRUs, microcontrollers and AI stuff
Sure, but these products are made to last long; deploying 32-bit applications in 2019 and hoping they’ll still work fine 5 years ahead is problematic considering how forgotten will the 32-bit landscape be in 5 years. Stability bugs and security issues affecting ARMv7 will be much less looked at by then, once no more Android device ships with a 32-bit OS and even the RPi will have issued the RPi-X for its 10th edition and finally migrated to 64-bit (which is far less certain :-)).
>hoping they’ll still work fine 5 years ahead is problematic considering
>how forgotten will the 32-bit landscape be in 5 years.
I’m not sure why you think this to be honest. M68K has been literally dead for a very very long
time now but you can still go and get the latest GCC for it.
>Stability bugs and security issues affecting ARMv7 will be much less looked at by then,
I think we’ll be seeing embedded vendors supporting ARMv7 for a long time to come. There is a place for 32bit cores and if RISC-V (or something else) really happens I think we’ll see licenses for the older ARM cores becoming very very cheap and attractive for every e-waste product your can think of.
> once no more Android device ships with a 32-bit OS
Android is hardly the be all and end all. If anything is going to be long gone in 5 years it will be Android itself.
I was speaking in terms of relevance only. Currently when a deep system vulnerability is reported which requires a per-architecture fix or work-around, like a ptrace bug, a race in exception handling, or anything (IMHO) less important like spectre/meltdown, x86 and ARM are the two first platforms to be fixed because they are important enough to assign resources there. But you make good points otherwise (including regarding android, though I’m not sure it’s for good considering that changes are driven by user friendliness, which is generally opposite to design quality :-))
A good chunk of ARMv7 stuff ships with kernels that came out way before spectre etc were discovered and will never get the fixes backported. We don’t need to wait 5 years for the ARMv7 ecosystem to be a security nightmare.
Board seems like a bone-ification of their X15 from a while back (2015; AM5729 vs 5728 on the X15). At half price, too.
X15 was their ‘super-beagle’ upscale project back then, but price was way too steep. Seems that has translated to the bone version as well.
The price is just insane for outdated technology. Not difficult to see why many will see numerous lucrative applications using RPi3B+ or RPi4
Which RPI has multiple DSP/MCU cores, a wide range of industrial and embedded interfaces, GPMC etc?
You’re killing Jerry’s dreams. He really believed RPi was made for industrial-grade products, and now a world falls apart 🙂
well, the board is nice, not too bad it’s 32 bit Cortex-A15, for me a system programming enthusiast the decent documentation is very good, but the price… pine makes a fully featured laptop for 200$, why this beagle should be so pricey? it’s like apple when the half of the price is for the bitten apple logo on the back, right? 😀
The reason why this is using an (arguably) outdated AM57xx Sitara device is TI’s “TI Deep Learning (TIDL)” suite of CNN/DNN packages, which are optimized for that product line. This should have been mentioned up-front in this article. [Note: TIDL is mentioned briefly in the article, toward the bottom.] See here for more:
https://training.ti.com/texas-instruments-deep-learning-tidl-overview