How the most popular Edge AI processors on the market compare Below we’ve compared the most popular AI frameworks presently on the market. The list is not exhaustive and we aim to keep this updated*

Processor Feature Arm GreenWaves Hailo Kneron Kneron NXP ST Syntiant Syntiant TI Sitara XMOS ETHOS-N57 GAP8 Hailo-8 KL520 KL720 Jetson Nano i.MX 8M Plus STM32F769xx NDP100 NDP101 AM5749 XCORE.AI

IP Block SoC - NPU SoC SoC - SoC - Single board SoC - MCU Ultra compact Compact SoC - SoC - Applications Applications Applications Development Kit Applications neural decision neural decision Applications Applications Description processor processor with processor with processor with processor processor processor processor with dedicated NPU dedicated NPU dedicated NPU dedicated NPU and GPU Mid range Application Primarily General General Application Smart Person Presence Voice Voice Automated General IoT devices include: Focused toward smart home connected include: Applications, Detection, Data applications, applications, sorting application balancing People counting, automotive applications devices Image Industrial IoT mining, Facial/ with support with support equipment, requiring AI performance, Road monitoring, sector, including: including: including: classification, voice recognition, for wake-word for wake-word Optical including: cost and battery consumer advanced driver Smart door locks, high-end IP Object detection, predictive detection. detection. inspection, Audio interface, life. Applications robotics, gesture assistance Drones, Smart Cams, Smart TVs, Segmentation maintenance, Applications Applications Vision computer, Presence include: recognition, systems (ADAS) doorbells, IP AI glasses and and Speech financial fraud include: Mobile include: Mobile Code readers, detection, Mainstream Face detection, and autonomous cameras, Robot headsets, and processing detection phones, Ear buds phones, Smart Industrial robots, Person Typical phones and Autonomous driving vacuums AIoT Gateways and hearing watches, IoT Logistics robots, identification, Application(s) digital TVs drone, Key word applications aids, Bluetooth endpoints, Smart Currency Actuation, Voice identification, headsets, Smart home, Remote counters, ATMs, interfaces, Surveillance watches, IoT Controls Patient monitors, Communications camera endpoints, Building and control Remote controls, automation, Smart speakers Industrial transport, Space, avionics & defence

Coprocessor Host Coprocessor Host Host Host (Standalone Host Host Coprocessor Coprocessor Host Host System Role (IP Block) development (Primary) board)

NPU (8 Cores) • 8 x RISC-V Proprietary • NPU • NPU (700MHz) • GPU • NPU (Vivante • ARM ARM Cortex-M0 ARM Cortex-M0 • 2 x Cortex-A15 • NPU cores structure-driven • Dual Arm • Cadence DSP (28-core VIP8000) Cortex-M7 • C66x VLIW • STM32M7 • Convolution data flow Cortex M4 (500 MHz) Maxwell • 4 x Arm • FPU DSP Architecture architecture @920MHz) Neural Network • Arm Cortex M4 Cortex-A53 • 2 × Dual ARM based on 16 Accelerator (400MHz) • 4 x Arm A57 (1.8 GHz) Cortex-M4 configurable (1.43 GHz) 'logical cores Processing • Arm Cortex-M7 • GPU (Dual Resource (800MHz) Core SGX544) • DSP (Cadence • 2 × EVE Tensilica HiFi 4) Analytic • GPU Processor (GC7000UL)

512 kB SRAM 512 kB SRAM Embedded Embedded Embedded 4 GB LPDDR4 LPDDR4/DDR4/ 512 kB SRAM Embedded Embedded 512 kB SRAM 1 MB SRAM RAM - LPDDR2 - 128 MB LPDDR3 (64-bit) DDR3L (16/32-bit) - 112 kB SRAM - 112 kB SRAM (2 x 512 kB Modules) Arm GreenWaves Hailo Kneron Kneron Nvidia NXP ST Syntiant Syntiant TI Sitara XMOS ETHOS-N57 GAP8 Hailo-8 KL520 KL720 Jetson Nano i.MX 8M Plus STM32F769xx NDP100 NDP101 AM5749 XCORE.AI 2 TOPS 22.65 GOPS 26 TOPS 0.28 TOPS (0.56 1.5 TOPS (0.9 472 GFLOPs 2.3 TOPS 1082 CoreMark 3,200 MIPS, (235 GOPS/W) (2.8TSOP/W) TOPS/W) TOPS/W) /462 DMIPS at 51.2 GMAC, and Performance 216 MHz 1,600 MFLOPS (7 CoreMark/ mW)

Several Several Several Several Several Several Several Several Several Several Caffe, Keras, TensorFlow Lite including: including: including: including: including: including: including: Caffe, including: including: including: TensorFlow, AI TensorFlow, TensorFlow TensorFlow and ONNX, ONNX, TensorFlow, TensorFlow, Caffe, Keras, TensorFlow TensorFlow TensorFlow Frameworks TensorFlow Lite, ONNX TensorFlow, TensorFlow, PyTorch, Caffe, TensorFlow Lite, TensorFlow Lite Lite, GluonCV, Caffe2, PyTorch, Keras, Caffe Keras, Caffe Keras, Darknet, and ONNX and ONNX MXNet, PyTorch, MXNet, ONNX MXNet ONNX

N/A SPI (Slave) PCIe SPI microSD HDMI 2.0 Subject to SPI Slave SPI Master USB 3.0 USB 2.0 Phy 2 x I2S (Rx-Only) RGMII UART 2 x MIPI CSI-2 (12 MIPI DSI device 2 x PDM SPI Slave USB 2.0 MIPI CSI (8 bit) 2 x MIPI CSI I2S lanes) LVDS (4 or 8 variant: Microphone I/P 2 x PDM 2 x PCIe 128 x GPIO 2 x I2C SDIO I2C Gb Ethernet Lane) 2 x USB 2.0 inputs microphone 4 x MMC/SD LPDDR DRAM UART USB 2.0 USB 2.0 M.2 18 x I2S 2 x SDIO inputs SATA 32 x GPIO I2C USB 3.0 HDMI S/PDIF UART 8x GPIO HDMI 1.4a 16 x PWM I2S PWM 4 x USB 3.0 8 x PDM Quad SPI 10 x UART Interfaces GPIO 2 x MIPI CSI I2C Microphone I/P I2C 4 x SPI UART SPI I2S 2 x USB 2.0/3.0 I2S 2 x CAN C2C DVP VI SPI 2 x Gb Ethernet CAN 2.0 5 x I2C DVP VO UART 2 x CAN HDMI GPIO PCIe Ethernet 4 x UART MDIO MIPI DSI SPDIF

Power <5W 0.5W 1.2W 10W TDP < 200 μW < 200 μW Consumption

N/A 88 aQFN 324 BGA 9mm x 9mm 260-pin edge 15mm x 15mm 100/144/176/208 12 WLBGA 32 QFN 760-pin BGA (7mm x 7mm) (15mm x 15mm) 11mm x 11mm connector LQFP (1.4mm x 1.8mm) (5mm x 5mm) (23mm × 23mm) Package (69 mm x 45 100/216 TFBGA mm) 176 UFBGA 180 WLCSP

Small package Very good Extensive I/O • Large Extensive I/O • RISC • Ultra low • Increased • High Low power power efficiency developer architecture power (< I/O when performance consumption community • Extensive I/O 140 μW compared to • TI open • Excellent • ST support recognising NDP100 community Differentiating support forum words) • Small package Features • • Very small processor package • GPU based AI acceleration • Extensive I/O

Contact us: Stuart Griffin, Founding Director & Technologist Adam Hoare, Business Development [email protected] [email protected] consult.red/edge-ai