Industrial AI Product of the Year
Company Name: Winbond Corporation
Product Name: CUBE/CUBE-Lite
Supporting Statement:
Winbond’s CUBE architecture removes one of the most significant engineering challenges limiting the next generation of edge AI systems: memory bandwidth efficiency.
AI performance at the edge is often determined by the efficiency with which data can move between memory and compute. In many systems, the memory interface has become the primary limitation on overall throughput. Once bandwidth, power consumption, routing complexity, and thermal limits are reached, adding more compute delivers limited real-world benefit. This is particularly challenging in compact edge devices where board space, energy efficiency, and thermal control are tightly balanced.
Conventional approaches to increasing bandwidth introduce trade-offs that are difficult to sustain in edge environments. Wider interfaces increase pin count, package complexity, and PCB routing density, while higher signalling speeds increase switching losses and thermal load. Existing high-bandwidth memory approaches designed for data centre infrastructure are often unsuitable for compact, power-sensitive edge AI systems.
Winbond developed CUBE (Customised Ultra-Bandwidth Elements) specifically for edge AI environments where bandwidth, power efficiency, thermal behaviour, and integration density must all be balanced simultaneously.
CUBE is an advanced memory architecture engineered for AI workloads requiring high throughput within strict power and thermal requirements. It supports both Chip-on-Wafer (CoW) and Wafer-on-Wafer (WoW) front-end integration and is compatible with advanced 3D packaging technologies, including Cap-interposer integration for enhanced power delivery performance.
At the device level, CUBE combines area-efficient Through-Silicon Via (TSV) structures with advanced die-to-die interconnect technologies including μbump and hybrid bonding. This enables significantly higher I/O density, improved transfer efficiency, and reduced interconnect overhead compared with conventional discrete memory approaches. Rather than scaling bandwidth through wider external interfaces and higher signalling overhead, CUBE increases local data throughput through vertically integrated memory architecture and shorter interconnect paths.
The architecture delivers up to 256GB/s bandwidth per die, while a 4Hi stack can scale to 1TB/s total bandwidth with energy efficiency below 1pJ/bit. This combination of throughput and efficiency enables AI edge systems to process larger datasets locally without introducing the thermal and power penalties typically associated with higher-bandwidth memory subsystems.
CUBE also provides flexible memory scaling, supporting densities ranging from 1GB to 2GB per die. A 4Hi configuration provides 4GB to 8GB total memory capacity, allowing system designers to align memory resources with AI model complexity and workload requirements.
Its compact footprint is another important technical advantage. By stacking compute directly above memory, CUBE reduces board-level routing complexity and minimises the physical distance data must travel between processor and memory. This improves transfer efficiency while reducing overall system footprint, making the architecture particularly well-suited to physical AI systems, including humanoid robotics, smart vision platforms, autonomous drones, AR/VR devices, and industrial edge computing systems.
Humanoid robotics represents one of the clearest examples of where CUBE delivers practical system-level benefits. These platforms combine real-time sensor fusion, local AI inference, machine vision, motion control, and low-latency decision-making within highly demanding thermal and mechanical envelopes. Sustained access to large volumes of data is essential, yet conventional memory architectures often introduce excessive power consumption, routing complexity, or thermal overhead. CUBE enables significantly higher local data throughput without relying on the larger, more power-intensive memory architectures commonly associated with data centre AI processing.
Alongside CUBE, Winbond is developing CUBE-Lite to support AI endpoint devices where power consumption, integration simplicity, and compact form factors are more critical than peak bandwidth.
While CUBE is designed for bandwidth-intensive AI systems that require sustained high-throughput data movement, CUBE-Lite is optimised for TinyML and always-on edge AI applications operating within tight power and thermal limits. Many low-power AI devices, including AI glasses, wearable systems, smart cameras, and intelligent sensors, cannot support the thermal overhead, package complexity, or energy demands associated with conventional high-bandwidth memory architectures.
CUBE-Lite applies the same core architectural principles developed for CUBE, including closer memory-to-compute integration and reduced interconnect overhead, but tailors the implementation for low-power AI SoCs performing continuous local inference. By eliminating the complex LPDDR PHY layer, CUBE-Lite simplifies SoC controller integration, reduces chip die area, and accelerates development timelines for edge AI device manufacturers.
This enables efficient on-device AI processing while reducing board space, simplifying integration, and lowering overall system power consumption. CUBE-Lite supports a broad range of scalable AIoT deployments, including smart cameras, battery-powered wearable devices, TinyML applications, and emerging PicoClaw-powered Tiny Agentic AI platforms. It demonstrates that advanced AI functionality can be deployed efficiently in compact edge systems without relying on large, power-intensive memory architectures.
Winbond has also strengthened the wider edge AI platform through integration with its TrustME® W77Q Secure Flash family. These devices provide hardware-based protection for firmware and AI model storage. Security features include secure boot, authenticated over-the-air updates, rollback protection, and root-of-trust validation. Support for LMS-OTS signatures provides a path towards post-quantum readiness, while alignment with Common Criteria EAL2+, SESIP Level 2, ISO 21434, and FIPS 140-3 (CAVP in progress) demonstrates compatibility with current security requirements for connected intelligent systems.
Together, CUBE, CUBE-Lite, and W77Q reduce multiple interconnected challenges within edge AI infrastructure: bandwidth scalability, power efficiency, thermal management, integration density, and firmware integrity.
Beyond the architecture itself, Winbond is actively collaborating with foundries, OSAT partners, and ASIC developers to build a broader ecosystem supporting next-generation edge AI integration. This ecosystem approach strengthens manufacturability, interoperability, and long-term deployment readiness across multiple intelligent edge applications.
CUBE enables AI edge systems to scale beyond the bandwidth and integration limitations restricting conventional memory architectures. By combining high-bandwidth memory architecture, advanced packaging integration, and efficient local data movement, CUBE enables practical AI deployment in compact edge systems where conventional approaches struggle to sustain throughput, thermal efficiency, and integration density.
Entry ID: 7521
Supporting Documents
CUBE_20260212.pdf, Memory-Architecture-for-Intelligent-Edge-Systems_EN_0427.docx