The new SL 1680 chip in the Astra lineup from Synaptics
Synaptics launched the Astra “AI-native” embedded computing platform at the Embedded World 2024 conference last week. Astra extends the company’s long history of edge innovation, focused on interfacing computers with humans and devices. Synaptics pioneered capacitive touch sensing in the 1980s and 1990s, using simple neural networks to recognize patterns and gestures on touch sensors. In 2017, Synaptics began aggressively expanding into the IoT space, culminating in Astra—a general-purpose AI platform designed specifically for edge applications.
Astra enters the market at the right time with the right features for rapid evaluation and adoption. While planet-scale AI grabs the high-tech headlines, a quiet revolution is underway at the edge of the network. AI is a game-changer for embedded computing systems because many use cases require high-speed analysis of diverse data types across multiple data streams. Machine learning algorithms are ideally suited for these kinds of pattern-matching tasks, while lower-tech alternatives, such as rule-based coding, are impractical at scale. Developers are highly motivated to run AI inference algorithms on small, embedded processors to generate actionable results at the point of data collection without the cost, complexity, delay and privacy risk of sending raw data to cloud services.
Edge AI applications require a new generation of high-performance, accelerated platforms explicitly designed for embedded computing with appropriate and carefully tuned performance, power, networking, I/O and cost profiles. Embedded AI platforms need five fundamental components: accelerated SoC hardware, a complete kit of platform software, standards-based wireless connectivity, a robust AI framework and a thriving partner ecosystem. The following diagram from Synaptics captures this vision neatly, so let’s see how Astra measures up.
Synaptics touts its Astra platform as covering all aspects of AI computing for IoT.
SoC Hardware (Compute Solutions)
At launch, the Astra silicon portfolio comprises three SoCs with Arm Cortex-A processors, each with a GPU and a full complement of I/O. As expected from an IoT company, the SoC configurations reflect a savvy understanding of edge platform requirements for AI use cases ranging from basic functionality with modest GPU acceleration to high-performance applications requiring NPU and GPU acceleration. All SL-Series SoCs also have hardware-accelerated security, video, graphics and audio.
The following tables summarize SL-Series SoC configurations and use case examples from Synaptics product briefs and other materials.
The different models available in Synaptics’ SL-Series of chips.
End uses for each of the SL-Series of chips from Synaptics
The SL-Series processors are available now. Later this year, Synaptics plans to expand the Astra product line with the SR-Series of microcontroller-based SoCs. These power-optimized AI-native chips use the same AI frameworks as the SL-Series, enabling lower-cost and battery-powered applications with the same AI framework.
For prototyping, Synaptics offers the Astra Machina Foundation Series evaluation kit. Its modular design has an I/O motherboard that supports swappable core modules for each SL-Series chip and daughter cards for wireless connectivity, debug and I/O expansion.
Platform Software (Unified Software Experience)
Astra’s GitHub repository has a complete set of open-source software for the platform, including Yocto Linux, the Astra SDK, the SyNAP toolchain, documentation and developer guides. Unified software means all the platform software and tools are in one place—and open source.
For new customers, the documentation includes straightforward procedures for getting started with Astra and the Machina Foundation evaluation kit. After configuring the hardware, the preloaded Linux distribution boots from onboard eMMC flash memory. Developers may then download a pre-built system image on the device, install a toolchain package for development and begin building applications. Developers can run AI applications immediately by compiling pre-converted neural networks or by converting and compiling custom neural networks with the SyNAP toolkit (described in the following section).
AI Framework
Synaptics has a long history of applying AI techniques to multimodal AI applications. The company introduced the Synaptics Neural Network Acceleration and Processing toolchain five years ago to support edge AI on AudioSmart and VideoSmart high-performance multimedia chips. It’s the core of Astra’s AI framework, converting neural networks from original representations (e.g., TensorFlow Lite and ONNX) into executable representations optimized for the target hardware, including specific NPUs and GPUs. The conversion may optionally involve model quantization, reducing model weight precision in favor of more efficient computation.
At Embedded World 2024, the Astra team demonstrated the process by downloading YOLO models (face detection) in both TensorFlow Lite and ONNX formats, running them as-is on the Astra Machina Eval Kit and using SyNAP to achieve two orders of magnitude of improvement in latency. SyNAP is a mature, open, accessible edge AI developer toolchain interfaced with industry-standard frameworks.
Wireless Connectivity
The Synaptics wireless portfolio offers mature solutions for Wi-Fi (including 6E), Bluetooth, Thread and Zigbee. For prototyping, the M.2 slot on the Astra Machina I/O board currently accepts three Wi-Fi and Bluetooth combination modules.
Moor Insights & Strategy recommends using Matter over Wi-Fi or Thread for consumer applications. Synaptics offers complete hardware and software implementations for Matter and Thread and is a member of both organizations. The Machina platform currently supports Matter over Wi-Fi. However, Matter over Thread must wait until Machina supports a combo module based on the SYN4381 or SYN4382 Wi-Fi, Bluetooth and Thread (802.15.4) chips. These parts are readily available, so we expect Machina to support Thread very soon.
Partner Ecosystem
Synaptics has been a supplier to many large OEMs and system platforms for decades. We expect the Synaptics ecosystem to evaluate Astra quickly, and the company can build upon longstanding relationships as Astra gains traction. Synaptics also has extensive supply chain partnerships, enabling production to scale with demand.
At Embedded World 2024, Synaptics showcased initial innovation partnerships, including these AI-enabled applications and integrated compute and connectivity solutions.
- embedUR — Image segmentation on an NPU-accelerated MCU for vision applications
- FocusAI – Enterprise SaaS with edge AI people detection and recognition running on Astra
- Darwin Edge — Vital signs monitoring for healthcare
- Fraunhofer — upHear real-time super-wideband audio enhancement
- Matter demo — AMPAK system-on-module with the SL1620 SoC and SYN43711 Wi-Fi/BT combo chip
Summing Up
Edge AI has transitioned from a “nice to have” experimental feature to an essential requirement for embedded platform architecture. IoT is a primary source of truth for training and inference, so AI’s explosive growth accelerates demand for smarter embedded devices. Although Astra is a new AI platform, it is surprisingly mature because it builds upon Synaptics’ long history with neural networks and sensor devices, extensive connectivity portfolio and broad partner base. We expect rapid adoption, particularly for multimodal AI use cases where Synaptics has deep expertise.
Source link
