Mosaic-SoC is building a perception-focused chip designed to bring real-time spatial awareness to the next generation of smart devices. Its architecture targets ultra-low-power environments such as wearables, mobile systems, and robotics, where continuous sensing must operate within strict energy and size limits. By combining vision processing with on-device AI, Mosaic aims to enable devices to interpret their surroundings directly, without relying on cloud processing, opening new possibilities for persistent, context-aware interaction across a wide range of applications.
Mosaic-SoC targets the intersection of AI and vision with a perception-first architecture designed for always-on spatial intelligence. The company develops a compact SoC that integrates general-purpose compute, on-chip memory, and dedicated accelerators into a tightly coupled system. This design supports real-time processing of sensor data in power-constrained environments such as AR devices, mobile platforms, and robotics.
The premise behind Mosaic’s approach reflects a broader shift in computing. Devices increasingly rely on cameras and sensors to interpret their surroundings, which raises the need for local processing that avoids latency and power costs associated with cloud inference. Mosaic addresses this requirement by focusing on direct, on-device processing of spatial data, allowing systems to map and respond to their environment in real time.
The chip operates primarily as a vision processor, with AI inference layered on top of spatial processing tasks. This positioning aligns with other edge AI vendors that integrate inference acceleration within a broader sensor-processing pipeline. Mosaic has not disclosed a formal NPU block or TOPS rating, and the company remains in an early stage of development. Its current status is pre-production, and detailed architectural disclosures such as block diagrams or datasheets have not been released.
Mosaic’s design centers on a memory architecture that supports concurrent access from multiple accelerators and compute units. Traditional SoCs often rely on shared pathways that serialize memory access, which limits throughput when multiple subsystems operate simultaneously. Mosaic addresses this constraint by enabling parallel read-and-write operations across accelerators, which allows continuous processing of data streams such as video, motion sensing, and audio inputs.
This capability supports workloads such as SLAM, eye tracking, and hand tracking, all of which require rapid processing of high-volume sensor data. These applications place sustained demands on bandwidth and latency, particularly in systems that must operate continuously within tight energy budgets. By allowing multiple subsystems to operate concurrently without memory contention, Mosaic aims to sustain throughput while maintaining low power consumption.
The company’s technical foundation traces back to the PULP platform, an open multi-core RISC-V architecture developed at ETH Zürich and the University of Bologna. Mosaic’s founders, Alfio Di Mauro and Moritz Scherer, both trained at ETH Zürich and worked on low-power processor design for edge AI workloads. Their background informs the company’s architectural direction, which extends PULP concepts into a commercial SoC optimized for perception-driven applications. Professor Luca Benini, a key figure behind PULP, serves as an adviser.
Mosaic-SoC was incorporated in April 2024 and currently operates with a small team. The company has raised limited funding to date and has initiated efforts to secure additional capital for next-generation chip development. It has also engaged with industry players including Meta, NXP, and GlobalFoundries in prior design work, which provides exposure to advanced semiconductor development and manufacturing processes.
For classification purposes, Mosaic fits within edge AI and vision processing categories. Its architecture aligns with embedded SoC designs that combine compute, memory, and accelerators for localized processing. Without published performance metrics, the chip cannot yet be benchmarked against established AI processors, but its focus on perception and spatial intelligence differentiates it from general-purpose inference accelerators.
What do we think?
Mosaic-SoC addresses a clear requirement for low-power spatial intelligence at the edge. Its architecture targets real-time perception rather than general-purpose AI acceleration, which aligns with emerging device needs. The company remains early, with limited disclosures and no production silicon. The key question is execution: whether Mosaic can translate its architecture into a scalable product within a competitive edge AI landscape.
Mosaic-SoC reflects an inflection point in AI, where perception and spatial awareness move directly into end devices, rather than relying on centralized compute. This shift toward localized intelligence changes system design priorities, emphasizing power efficiency, latency, and continuous sensing. If devices widely adopt perception-first architectures, this inflection point could redefine how AI integrates into everyday systems, expanding the role of edge processing and reducing dependence on cloud-based inference across consumer and industrial applications.
LIKE IT? THINK YOUR FRIENDS AND ASSOCIATES MIGHT? PLEASE SEND IT TO THEM WITH OUR BEST WISHES.