Foxconn, Socionext, and Hailo are launching a fanless “BOXiedge” AI edge server with Socionext’s Linux-driven SynQuacer SC2A11 SoC with 24x Cortex-A53 cores. The system also features an up to 26-TOPS Hailo-8 NPU for analyzing 20 streaming cam feeds in real time.

In January, Foxconn, Socionext, and Network Optix announced a BOXiedge AI video analysis solution that combined a Foxconn edge computer equipped with an unnamed AI acceleration card and Socionext’s SC2A11 Arm server SoC. The BOXiedge also incorporated Network Optix’s Ubuntu compatible Nx Witness v4.0 video processing and analytics software. Now Foxconn and Socionext have followed up with a new BOXiedge model that omits the Nx Witness v4.0 stack and replaces the AI card with an up to 26 TOPS Hailo-8 deep learning processor.

BOXiedge v2 (left) and Hailo8 SDK architecture
(click images to enlarge)

The new BOXiedge v2 “local VMS server,” which we read about on VentureBeat, is designed for AI-enabled industrial internet of things, smart cities, and smart medical applications that require a compact footprint, low latency, and low power consumption. Examples include tumor detection, traffic monitoring, robot navigation, product inspection, and quality assurance. The system can perform real-time image classification, detection, pose estimation, and other tasks on footage from up to 20 cameras while running at 35W.

The BOXiedge product page was not launching at publication time, which may be on purpose. Judging from the image, however, the fanless, desktop system appears to have 2x GbE, 2x USB 3.0, and a console port. We will check back later to see if we can find some more details.

 
Hailo-8

Hailo’s 17 x 17mm Hailo-8 AI chip performs at 3 TOPS per Watt, and can achieve up to 26 TOPS with higher power consumption. Its benchmarks show the Hailo-8 vastly outperforming Nvidia’s AGX Xavier and Google’s Edge TPU on a TOPS per watt basis running AI semantic segmentation and object detection applications including ResNet-50.

Hailo-8

The Hailo-8 uses a “proprietary novel structure-driven” Dataflow architecture that differs from the Van Neumann architecture used on most neural processors. The architecture achieves low-power memory access by implementing a distributed memory fabric combined with purpose-made pipeline elements. A Dataflow-oriented interconnect adapts according to the structure of the neural network to enable high resource utilization, says Hailo.

The Hailo-8 chip is tightly integrated with an SDK that offers scalable toolchain including model translation from industry standard frameworks like ONNX and TensorFlow. According to VentureBeat, Foxconn is one of Hailo’s first publicly disclosed customers after NEC and ABB Technology.

 
SynQuacer SC2A11

The BOXiedge announcement made no mention of Linux, but Socionext’s SynQuacer SC2A11 SoC supports Linux. Designed for parallel processing of video data, the SynQuacer SC2A11 combines 24x, 1GHz Cortex-A53 cores connected via a 4MB L3 equipped cache coherent interconnect. You can combine up to 64x multiple SC2A11 chips with a combined 1,536 cores running in parallel.

The 30 x 30mm BGA packaged SC2A11 frees up computer resources due to its modest footprint, which enables the OS to “directly assign processes for its cores without employing virtualization,” says Socionext. The SoC incorporates 32KB/32KB L1 per core and 256KB L2 cache for every two cores. The SC2A11 features dual 4-lane PCIe Gen2 interfaces and 2133Mbps DDR4 support, including ECC RAM. It also provides 2-ch. Gigabit Ethernet support along with an IPSec Network Offload Engine.

SynQuacer SC2A11 block diagram (left) and Socionext Developerbox
(click images to enlarge)

The SynQuacer SC2A11 has appeared on a 96Boards Enterprise Edition based Socionext Developerbox system, which is built around a Micro-ATX board. Developed with Linaro and Gigabyte, the system runs a Linaro version of Ubuntu and is referred to as the SynQuacer E-Series by Socionext.

Back in 2018, Phoronix ran some benchmarks and found the Developerbox to be somewhat wanting considering its $1,200 price ($462 for the mainboard only). However, the story noted that its benchmarks were not designed to exploit the SC2A11’s strengths in running all 24 cores simultaneously. Last year, CNXSoft reported on a V-Raptor PCIe board based on the SC2A11 from XSLAB.

Earlier Socionext SoCs include the Linux-powered Socionext MB86S71 SoC, which has a pair each of Cortex-A15 and -A7 cores. The MB86S71 appeared on the 96Boards CE compatible Fujitsu F-Cue SBC.

 
Further information

No pricing or availability information was provided for the BOXiedge. More information may be found in the joint announcement and this Embedded Computing Design blog post from Socionext VP Rajinder Cheema. More may eventually appear at this currently unresponsive BOXiedge v2 page. It is unclear whether the Nx Witness v4.0 equipped BOXiedge is still available or ever reached market.