Since it was first coined in the 1950s, artificial intelligence (AI) and Machine Learning (ML) have colonized a substantial share of advanced machines. This happened as AI computing, specifically ML, enables a better quality of life with the technology branching into both homes (personal assistant robots) and transportation (self-driven vehicles), and even into persuasive advertising (image-recognition). The list of AI applications seems infinite, bolstered by smarter, more accessible, and more accurate systems.
Both AI and ML computing are synthesized inside a data warehouse, or remotely in a Cloud, or at the “edge.” Edge computing is ideal when an already low latency and real-time response plunge to critical levels. Several SBC platforms are available for AI and ML computation. This article will compare BeagleBone AI and Ultra96 boards. Both play a key role in AI and ML.
BeagleBone AI is one of the quickest vehicles to embedded AI at the edge. This super flexible and fast AI is the end product of multiple years’ research in open hardware single-board Linux computers. You can use it to automate your shop floor, home, office, or lab.
The BeagleBone AI draws its strength from the 1.5GHz, dual-core Cortex-A15 Texas Instruments Sitara AM5729. Even though the CPU lags behind its competitors in the Linux hacker board arena, the key selling points are several coprocessors, led by the TI 4x embedded-vision-engine (EVE) neural processing cores with the AI capabilities of the SoC.
Such capabilities help easier exploration of how AI and ML can be integrated into everyday life. The AM5729 integrates 2x TI C66x digital signal processors (DSPs) with 2x Cortex-M4 MCUs plus the EVE cores. Similar to the previous SoC, the AM5729 is fortified with a 4x programmable, real-time PRU-ICS (Programmable Real-Time Unit and Industrial Communication SubSystem) cores. The PowerVR SGX544 3D GPU and Vivante GC320 Core 2D accelerator are identical as in the previous iterations. The pre-installed software tools and optimized TI Deep Learning (TIDL) software framework support ML. Downloading is simple- all you need to do is a few clicks- and the latest instances of edge inference applications and algorithms and applications will run.
This board is useful for everyday automation in the home, industrial, and commercial applications. Part of its high popularity can be ascribed to its competitive pricing.
BeagleBone AI Features
The BeagleBone AI runs on a TI Sitara AM5729 processor, with a 2x Cortex-A15 @ 1.5GHz with 2.5MB of L3 cache. The GPU is a PowerVR SGX544 3D armed with a Vivante GC320 Core 2D accelerator. Enhanced performance is pushed by 4x Embedded Vision Engines (EVEs) with Vision AccelerationPac architecture armed with 2x Cortex-M4 and 2x dual-core PRU (4x cores total). The AI has 1GB DDR3L (2x 512MB) and 16GB eMMC memory supplemented by a MicroSD slot.
The BeagleBone AI networks through a Gigabit Ethernet port funneling a Dual-band 802.11ac with Bluetooth 4.2 (AzureWave AW-CM256SM). The Media I/O consists of a 16-bit LCD interface and a Micro-HDMI 1.4a port at up to 1080p @ 30fps or 4K @ 15fps (H.264) with a Touchscreen controller. The other I/O are USB 3.0 Type-C dual-role port with power and I/O, USB 2.0 host port, 4x UART, 2x I2C, 2x SPI, and PRU I/O pins with a JTAG Serial debug header. Expansion is made possible by the BeagleBone Cape compatible expansion header (2x 46-pin). The other features are 5x LEDs, heatsink, antenna, cables, and fan support.
Power gets fed to the BeagleBone AI through USB Type-C or expansion pin input and TI TPS6590379 PMIC. The component has 7x step-down converters for 6A output plus 3A regulator for up to 9A output. There are power and reset buttons. It uses a Debian 9.9 Operating System.
Avnet’s Ultra96 is an inexpensive, easy-to-use platform based on the integrated Dual-core Arm Cortex-R5F real-time, multiprocessing system with programmable logic Xilinx Zynq UltraScale+ MPSoC. The hardware accelerates compute-intensive tasks.
Developers enjoy a commanding and exceptional environment to simplify ML. Avnet and Xilinx are keen to explore what types of MLs and edge AI designs can be developed by means of the newly introduced Ultra96 Development Board.
This is only one instance of the many excellent use cases for the Avnet Ultra96 when ML is implemented at the edge. The use of Zynq MPSoC makes this possible, as it offers both programmable logic and Arm processor cores. The heterogeneous device allows you to architect a solution that achieves the perfect balance between performance and power-both essential parameters- for edge centric applications. This fine balance between performance and power is accomplished by the use of programmable logic to accelerate the ML function.
Performance is better compared to a purely PS-based solution, and if properly implemented, it will bonus to significant power advantages too. To demonstrate its acceleration capabilities, Xilinx offered a Binary Neural Network (BNN) reference design for the ZCU102 and Ultra96. The list of edge application categories are inclusive of, but not restricted to, consumer electronics, autonomous vehicles, home or industrial IoT, drones or robots.
The Ultra96 boots from the provided PetaLinux pre-loaded Delkin Devices manufactured Industrial 16 GB MicroSD card. Designers connect to Ultra96 via a web server using integrated wireless access point capability. They can also connect through the PetaLinux desktop environment. The result can be viewed on the integrated Mini DisplayPort video output. Several application examples and numerous on-board development options are offered as reference.
Avnet released their 96Boards Click Mezzanine to offer engineers the advantage of expanded capabilities for the Ultra96 platform. This economically priced, add-on board allows designers to connect 96Boards low-speed (LS) header to MikroElektronika manufactured Click Boards. Sensors, relays, controls, and more can be latched to their MPSoC. Such Click Boards offer personalization for end applications, reducing any risk associated with designing on custom boards. The engineering resources are also efficiently applied across the complete design process.
Ultra96 board Features
The Ultra96 has several superior features like the Xilinx Zynq UltraScale+ MPSoC ZU3EG A484 ably supported by the Micron 2 GB (512M x32) LPDDR4 RAM. It is powered by the 16 GB microSD card pre-loaded with Embedded Linux OS and compatible with 802.11b/g/n Wi-Fi and Bluetooth 4.2 (provides both Classic & BLE). There are a Mini DisplayPort (MiniDP or mDP), 1x USB 3.0 Type Micro-B upstream (Device) port, and 2x USB 3.0 and 1x USB 2.0 Type-A downstream (Host) ports. Users will also find a 40-pin Low-speed expansion header and a 60-pin High-speed expansion header.
We compared them. What Do you think? What is your preference when it comes to Machine Learning: BeagleBone AI or the Ultra96-V2?