Embedded HPC and Embedded Deep Learning

---

Embedded and remote applications prohibit conventional HPC and deep learning solutions due to constraints on space (volume) and power consumption, and in some cases weight (collectively known as "SWaP", or size, weight, and power consumption). It's prohibitive to install a 300 W, 2-slot thick, 7 lb. GPU board (or equivalent Xeon Phi board) on a small form-factor motherboard such as pico-ITX, mini-ITX, PC104+, or similar.

For small Edge and IoT applications, a variety of small, low-power embedded targets have emerged, as listed below.

Embedded HPC and Embedded Deep Learning Targets

• Nvidia Jetson TX2 (Tegra Parker SoC, 256 CUDA cores)
• Intel Atom dual and quad-core CPUs (x5-E39xx CPUs)
• Movidius (Intel) Myriad ASIC (deep learning only)
• ARM Raspberry Pi (quad core, 1.2 GHz)

More info on the above HPC and deep learning embedded targets, including images and detailed descriptions, plus info on running compressed models on these targets including MobileNet and SqueezeNet, is on the SigDL Github page.

From pico-ITX and mini-ITX servers on up, coCPU accelerators can also solve these constraints, providing a dramatic increase in "HPC Density", expressed as Performance / Power / Volume / Weight. (GFlop/W/cm³/lb).

Embedded HPC and Embedded Deep Learning Applications

• Robot HPC, Robot Vision
• Drone HPC, Drone Video
• Satellite HPC
• Vehicle HPC and mobile HPC - truck, ship, military
• Remote Equipment
• Dense, private clouds
• Self-Driving Vehicles
• Harsh Environment HPC

coCPU Accelerator Features

• Low Size, Weight and Power (Low SWaP)
• Single-slot width, half-length PCIe card
• Accelerate Atom CPU based mini-ITX motherboards by 500x
• OpenMP programming interface (optional, requires C/C++ source)
• Supports Atom and ARM based small form-factor motherboards, including pico-ITX and mini-ITX

coCPU Accelerator Basic Specs

• DirectCore software support
• CIM software support
• Application specific example and demo software
• 0.64 Teraflops - expandable to 1.28 TFlops
• 50 W (100 W at 1.28 TFlops)
• thirty-two (32) TI C66x cores
• 1 GbE NIC (access to all cores)

Embedded HPC Motherboard Features

• Intel® Atom™ Processor x5-E39xx, 1.3 to 1.6 GHz
• 1x DDR2 SO-DIMM, Max. 2GB
• Supports CRT and LVDS (dual display)
• Onboard Dual PCIe Gigabit LAN
• 2x SATA, 8x USB 2.0, 4x COM, 1x IDE
• 1x PCI-E (x4), Watchdog timer, Digital I/O

Embedded HPC and Deep Learning Lab / Test Setup