Winbond Launches CUBE Architecture Bringing Ultra-High Bandwidth Memory to Edge AI

Winbond, a global leader in memory solutions, today announced a powerful memory-enabling technology that helps customers achieve affordable edge computing in mainstream applications. Winbond’s CUBE (semi-customized ultra-high bandwidth element) significantly optimizes memory technology to achieve performance for running generative AI in hybrid cloud and edge cloud applications.

CUBE enhances the performance of front-end 3D structures, such as chip-on-wr (CoW) and wafer-on-wafer (WoW), as well as back-end 2.5D/3D chip-on-Si-interposer substrates and fan-out solutions. CUBE is designed to meet the growing demand for edge AI computing devices, and can use 3D stacking technology and combined with heterogeneous bonding technology to provide high-bandwidth, low-power single-chip 256Mb to 8Gb memory. In addition, CUBE can also use 3D stacking technology to enhance bandwidth and reduce the power required for data transmission.

The launch of CUBE is an important step for Winbond to achieve cross-platform and seamless deployment. CUBE is suitable for advanced applications such as wearable devices, edge server devices, monitoring equipment, AS and collaboration.

Winbond said

The CUBE architecture transforms AI deployment, and we believe the integration of cloud and edge AI will take AI to the next level. We are unlocking new possibilities with the CUBE and are improving memory performance and optimizing costs for powerful edge AI devices.

CUBE's main features:

Save power

CUBE offers exceptional efficiency, consuming less than 1pJ/bit, ensuring extended runtime and optimized energy usage.

Superior performance

With bandwidth ranging from 32GB/s to 256GB/s, CUBE delivers performance gains well above industry standards.

Smaller size

CUBE has a smaller form factor. Currently based on the 20nm standard, it can provide 256Mb-8Gb capacity per chip, and the 16nm standard will be available in 2025. The introduction of through silicon vias (TSV) can further enhance performance, improve integrity, power integrity, reduce IO area with 9 um tch, and better heat dissipation (when CUBE is placed below and SoC is placed above).

High cost-effectiveness and high bandwidth

CUBE's IO speed can reach up to 2Gbps at 1K IO. When integrated with SoCs using mature processes such as 28nm and 22nm, CUBE can reach a bandwidth of 32GB/s to 256GB/s, which is equivalent to HBM2 bandwidth and 4 to 32 LP-4x 4266Mbps x16 IOs.

SoC chip size reduction

SoC (without TSV, on top) is stacked on CUBE (with TSV, on the bottom). If the TSV area loss is removed, the chip size may be smaller. It can bring more obvious cost advantages to edge AI devices.

Winbond said

CUBE can unleash the full potential of hybrid edge/cloud AI to improve system functionality, response time, and energy efficiency. Winbond's commitment to innovation and collaboration will help developers and enterprises jointly drive progress in various industries.

In addition, Winbond is actively working with partners to build a 3aaS platform that will further leverage the capabilities of CUBE. By combining CUBE with existing technologies, Winbond can provide the industry with cutting-edge solutions to enable enterprises to thrive in this critical era of AI-driven transformation.

Review editor: Liu Qing