On-device Processing Ramps Up With New Edge AI, End-point AI, & TinyML
While there are differences between edge AI, end-point AI, and TinyML, each of these fields is progressing real-time processing in its own ways. Below is a roundup of recent news in these three areas.
As its name suggests, edge artificial intelligence (AI) refers to the intersection of edge computing and artificial intelligence. This domain is related to (yet distinct from) end-point AI, the crossover between end-point computing and AI, and TinyML, the deployment of machine learning on microcontroller-class embedded devices.
The basic principles behind an edge AI system. Image courtesy of Digi-Key
Edge AI, end-point AI, and TinyML have some common objectives: deploy AI applications closer to where the data is collected, process information in real-time, operate independently of a network or in a closed system, reduce power consumption, and use limited computing capabilities. These objectives benefit real-time processing in automated vehicles, privacy in medical applications, and AI deployment in places that usually require cloud computing and data centers.
Below is a roundup of some of the latest industry news in the edge AI, end-point AI, and TinyML world.
Infineon Acquires Imagimob for Embedded AI
Germany’s largest semiconductor company, Infineon Technologies, recently announced the acquisition of Swedish AI company Imagimob.
Infineon is a spin-off of Siemens, with a product portfolio ranging from ASICs, microcontrollers, sensors, and small sensor transistors/diodes. Applications for these products span the Internet-of-Things (IoT), automotive, power systems, and beyond. Earlier this year, the company announced plans to build a chip manufacturing plant in Dresden with partial funding from the European Chips Act.
Founded in 2013, Imagimob is a company that focuses on automated machine learning (AutoML) tools specifically for embedded system deployment. The Imagimob AI development platform allows non-experts in ML to collect, annotate, manage, analyze, and process data; develop models; evaluate and verify multiple models; and package those models for deployment on embedded devices. The platform visualizes all of these steps and provides a user interface to manage models.
Screenshot of Imagimob’s AI platform. Image used courtesy of Imagimob
Infineon’s acquisition of Imagimob will complement its semiconductor portfolio with an AI platform that attempts to ease deploying machine learning on embedded devices.
Andes Technology Announces AndesAIRE
Taiwanese company Andes Technology recently launched its Andes AI Runs Everywhere (AndesAIRE) platform for edge AI and end-point AI applications.
AndesAIRE comprises several technologies and tools:
- Andes Deep Learning Accelerator (AnDLA) I350, a low-power deep learning accelerator that supports PyTorch, TensorFlow Lite, and ONNX
- Neural Net Software Development Kit (NN SDK) that includes NNPilot for optimizing neural networks, providing AnDLA drivers/runtimes, pruning/quantizing models, and generating the executables for deployment on AnDLA devices
AnDLA I350 block diagram. Image used courtesy of Andes Technology
Specs of the AnDLA I350 include:
- Int8 configurable MACs (32 to 4096)
- 8 TOPS at 1 GHz max performance
- 16 kb to 4 MB local memory
- Multi-dimension direct memory access
- 4 x 64-bit AXI bus interfaces
- CNN inference
- NN models for image/video and speech/voice/audio applications, including ResNet-8/50, Tiny YOLO v1/v2, LSTM, RNN, and GRU, among others
- Support for convolution, element-wise operations (add/sub/mul), various activations (ReLU, sigmoid, etc.), pooling, upsampling, concatenation, and normalization, among others
Andes Technology envisions the Andes ecosystem playing a role in advancing efficient AI/ML applications—for example, offloading tasks between the AnDLA I350 and AndesCore RISC-V CPU. The company is among the founding members of the RISC-V International Association.
Nuvoton Launches Heterogeneous Multi-core Microprocessor
Taiwanese Nuvoton Technology recently launched the NuMicro MA35D1 microprocessor series, designed for smart buildings, smart factories, and TinyML.
NuMicro MA35D1 microprocessor. Image used courtesy of Nuvoton
In particular, the MA35D1 series has been designed for high reliability and security and can satisfy IEC 62443 certification.
The technical specs are:
- 64-bit Arm Cortex-A35 cores
- A 180-MHz Arm Cortex-M4 core to support real-time control
- 800 MHz max clock speed
- Up to 512 MB of DDR3L SDRAM
- Support for USB, SD/eMMC, NAND, and SPI Flash
- LQFP and BGA packaging options
- Operational temperature range of -40°C to 105°C
- Integrated TFT LCD controller with up to 1920 x 1080 resolution and 60 fps
- JPEG and H.264 hardware decoders
- Support for Linux, RTOS, and OpenWRT operating systems
Supporting security includes:
- Nuvoton-developed Trusted Secure Island (TSI)
- Arm TrustZone technology
- Secure boot
- Tamper detection
- Encryption/decryption accelerators for AES, SHA, ECC, RSA, and SM2/3/4
- A true number generator (TRNG)
- Key storage memory
- One-time programmable (OTP) memory