The BeagleBone organisation has created a 64bit Linux single board computer with AI capability and other hardware accelerators. Eight Top/s of neural-network processing are accessible through Python libraries.
Called AI-64 it is built around Texas Instruments’ TDA4VM, which has 2GHz dual Arm Cortex-A72 cores, a C7x DSP core, and deep learning, vision and multimedia accelerators, according to Farnell, which is stocking the board and said that it is “designed for embedded applications and brings a complete artificial intelligence and machine learning system to developers”.
The 1GHz DSP core is rated at 80Gflop/s and 256Gop/s, and the 8Top/s of 8bit of neural network processing comes from TI’s ‘MMA’ (matrix multiply accelerator), also running at 1GHz.
Also on the die are two more DSPs – a pair of C66x floating-point VLIW cores, a microcontroller block with six Cortex-R5F cores and a PowerVR Rogue 8XE GE8430 3D GPU. Video acceleration includes 2x 1080p30 H.264 encode and 8x 1080p30 H.264/H.265 decode.
The numbers above are somewhat dependent on the version of the TDA4VM fitted – that part number represents a family. Farnell and BeagleBoard are vague on this – the data sheet calls up a TDA4VM88TGBALFR – and, for example, even TI is not definite on what the ‘x’ in ‘C7x’ represents.
Memory includes 4Gbyte ram, 16Gbyte eMMC flash and a MicroSD Card slot.
According to Farnell, interfaces include: M.2 E-key PCIe connector for WiFi-Bluetooth adaptors, USB 3.0 Type-C SuperSpeed interface for power input and data, two USB 3.0 Type-A SuperSpeed ports and Gigabit Ethernet.
On top of this there is Mini DisplayPort for a monitor, 4 lane CSI connector for a camera and a 4 lane DSI connector more displays.
Thee are boot, reset and power buttons, a power LED and five user-accessible LEDs.
An ikroBUS Shuttle header is provided forClick sensors and actuators, and “BeagleBone cape headers provide expansion possibilities, with hundreds of hardware examples and dozens of off-the-shelf embedded expansion options”, said Farnell. “A web browser, power source and network connection are all that are needed to start building embedded applications.”
It sees potential applications in autonomous robots, autonomous drones, smart buildings, smart factories; home security, retail automation, media servers, and machine vision.
This TI page (scroll down) has specs for the TDA4VM88TGBALFR, if that is the chip that is fitted.
Beagleboards download page for the schematic does not seem to be working at the time of typing, but there is a schematic here