Transform your smart devices with our cutting-edge ML accelerator IP blocks
Empower complex machine learning tasks with low power consumption. Why not upgrade to our smart edge processing solutions?
Machine learning accelerators for edge processing devices
CSEM’s custom silicon-proven machine learning accelerator solutions empower smart devices with advanced machine learning, offering exceptional performance and efficiency for ultra-low power applications.
CSEM’s energy-efficient machine learning (ML) hardware accelerators enable complex tasks on low-power edge devices. Our silicon-proven IP blocks support various ML algorithms and include efficiency-boosting tools, extracting meaningful data from sensor data such as audio, video, or medical signals. They achieve this while reducing computer time and extending battery life. Ideal for power-constrained edge applications, our solutions outperform traditional hardware and work with custom algorithms.
Energy-efficient AI/ML accelerators and IP solutions for edge-processing SoCs
We develop energy-efficient custom SoCs and hardware for low-power applications, focusing on high-performance ML accelerators for edge processing. Our expertise in silicon-proven design and proven IP blocks enhance smart devices by optimizing size, power consumption, and battery life.
CSEM's IP library offers a comprehensive suite of hardware IPs designed to create modular and flexible System-on-Chips (SoCs). These are tailored for end-to-end machine learning (ML) inference on compact, power-efficient systems.
We focus on the crucial components of the ML accelerator ecosystem:
- ML accelerators: for handling complex ML tasks efficiently, ensuring rapid data processing and real-time inference, essential for edge applications.
- Specialized memory and flexible interconnection systems: for optimized data flow and storage, reducing latency and power consumption critical to maintaining performance in power-constrained environments.
- RISC-V-based 32-bit processor core (icyflex-V): offering a unique balance between performance and energy efficiency, supporting a wide range of ML algorithms and applications.
- Various peripherals: enhancing the functionality and connectivity of SoCs, making them adaptable to different use cases and environments
We ensure seamless integration and optimized performance for edge AI/ML applications by offering and integrating these components. This allows us to deliver solutions that are not only powerful and efficient but also versatile and adaptable to the specific needs of our clients.
Comprehensive ML acceleration IP blocks
Parallel computing for dedicated ML tasks
- Binary decision tree (BDT) for ultra-low-power computer vision, such as presence detection, among other ML tasks. Based on early termination, a BDT minimizes computational effort due image content.
- Convolutional neural network (CNN) for complex computer vision tasks. Low-latency and low-power consumption are guaranteed by configurable parallel processing elements (PEs) and programmable precision. Supports several layer types, such as FC, Conv, DSConv, among others.
- Recurrent neural network (RNN) for ML analysis of time series signals as they capture long-term dependencies. Both LSTM and GRU layer types are supported.
Modular library for fast and simple integration into any system, covering the complete end-to-end processing datapath.
Available software/deployment stacks include an ML model compiler and a standard RISC-V toolchain, featuring a face detection example and support for various ML flows and formats, including TensorFlow, ONNX, PyTorch, and Caffe.
Discover detailed specifications in our IP Library Technical Factsheet: IP library for the acceleration of edge AI/ML.
Applications leveraging CSEM edge ML hardware
- Privacy-preserving occupancy detection
Detects occupancy in rooms, parking spots, etc., while preserving privacy. - Analysis of facial features for human-machine interfacing
Utilizes analysis of facial features for intuitive human-machine interaction. - Speech-based human-machine interfacing
Enables efficient voice recognition for smart devices.
- Condition monitoring of large ball bearings
Predictive maintenance through continuous monitoring. - Drone-based fire area assessment
Assists in fire fighting by assessing fire areas
Innovative solutions and System-on-Chip implementations
Our innovative edgeML solutions showcase the capabilities of advanced machine learning on low-power platforms. From emotion recognition on a credit-card-sized device powered by a coin cell battery to sophisticated System-on-Chip (SoC) designs such as the sub-mW Visage SoC and the multi-modal Fibonacci SoC, these solutions highlight the ability of our cutting-edge technology to deliver efficient, end-to-end ML processing for various applications. These solutions combine powerful machine learning accelerators with ultra-low power consumption, making them ideal for continuous (always-on) operation in smart devices.