Silicon Labs and Eta Compute Simplify Edge-ML Development
Artificial intelligence (AI) is the hot, trending buzzword these days. No technology-related conversation is complete without it being mentioned at least once. Such has been its impact on all areas of our lives. Coupled with its twin, machine learning (ML), it is revolutionizing everything from summarizing daily meetings to demystifying complex protein structures. These features come with considerable costs, including the enormous power it consumes, its dependency on data for training, and the inherent security risks associated with it as cyber criminals become more sophisticated. This is where AI/ML on the tiny edge comes into play.
The tiny edge is the deployment of computing and data processing capabilities on small, resource-constrained devices located closest to the data source. These devices, often sensors or microcontrollers with limited processing power and memory, perform operations locally rather than relying on cloud computing. The tiny edge enables running ML models on these small devices, allowing for real-time decision-making and data analysis directly, thereby cutting the costs associated with data transmission and cloud storage. This method is especially useful for applications that require low latency, low energy consumption, enhanced privacy and lower-bandwidth utilization.
Eta Compute is Empowering ML on the Tiny Edge
Eta Compute is a Silicon Valley startup with a mission of empowering machine learning on the tiny edge and innovation through solutions that overcome the gap between the fast-moving landscape of AI and the unique challenges of embedded systems. Fueled by a team of experts in ML, IoT, and systems design, Eta Compute understands the challenges of deploying ML models on resource-constrained edge devices.
Eta Compute’s Aptos is a revolutionary no-code software toolchain engineered for embedded inference. Designed to streamline edge-ML model development, Aptos accelerates the creation of efficient models tailored for low-power edge processors.
Aptos takes a new and unique approach to ML tools – it is itself based on ML techniques. These internal ML models allow Aptos to automatically learn the ML capabilities and performance characteristics of embedded processors. By harnessing these insights and abstracting away the hardware details, Aptos enables product developers to focus their domain and use-case expertise to create great edge-ML products without getting bogged down in the details of data science, chip strengths/weaknesses, neural network compilers, and resource constraints.
The Silicon Labs and Eta Compute Partnership Enables the Development of Advanced Edge ML Embedded Products
Silicon Labs is pleased to announce a partnership with Eta Compute that will empower product developers to seamlessly integrate advanced ML capabilities into their edge-ML embedded products.
Silicon Labs has long been known for innovations in low-power solutions for IoT, automotive, healthcare, industrial, and consumer products. Our solutions combine efficient microcontrollers and wireless connectivity with new capabilities for embedded inference. This technology leadership makes us a partner of choice for developers looking to add ML capabilities to their products. Yet, product developers face difficulties bringing the best ML techniques into their edge-ML embedded products. One challenge is resources – it’s difficult to find and retain the “unicorn” engineers with both expertise in ML and also an understanding of the tight requirements of low-power silicon (compared to the cloud environments an ML expert normally uses). This makes targeting and optimizing ML models to the constrained embedded system a huge challenge.
Another challenge is the successful adoption of new, rapidly evolving silicon with embedded ML capabilities. With the traditional approach, characterizing a new inference chip would entail a large effort by skilled ML/embedded practitioners to explore and understand in detail each new chip’s neural network capabilities in order to properly leverage its strengths and work around the inevitable embedded constraints such as the amount of memory or missing support of specific ML operations.
The result of these challenges with a traditional ML process is a lengthy product development cycle that is fraught with iterations. Many products remain stuck in proof-of-concept experiments and fail to make it to volume production.
The partnership with Eta Compute resulted in Aptos being trained on Silicon Labs solutions, including the EFR32xG24 and EFR32xG26. The support of these chips by Aptos was undertaken so that their mutual customers can use the platform to easily create highly optimal models running on Silicon Labs’ solutions for edge ML vision tasks like image recognition, categorization, counting, and human pose-detection.
Aptos automatically generates and characterizes optimal ML models targeted at the selected chips. The initial Silicon Labs chips supported by Aptos are the EFR32xG24 and EFR32xG26. Once the ML problem to be solved by Aptos is defined by specifying the objectives and constraints - such as model accuracy, latency, and power requirements - and using training and validation datasets, Aptos applies its knowledge of the respective Silicon Labs SoC and compilers to automatically generate models. These models are also automatically characterized by Aptos against the actual hardware development kits, so you can be assured that metrics reported by Aptos is indeed what you will achieve on your own Silicon Labs-based hardware.
Eta Compute developed Aptos to help companies realize the immense potential of Edge-AI/ML. This collaboration seeks to overcome the challenges of getting edge-ML products out of proof-of-concept experiments and into volume production, including addressing the huge gap between the worlds of ML and embedded software and the resource, schedule, and expertise constraints customers may experience.
We will have more details about using Aptos with Silicon Labs solutions in future blogs, and in the meantime you can see a live demo and guided tour of Aptos in action, or sign up for a free trial to generate your own optimized ML models targeting Silicon Labs solutions here.
