CES® 2027 NEWS
ROAD TO CES 2027Live now
ROAD TO CES · AN INTELITOYS EVENTJAN 6–9 · VEGAS
AI Devices

Winbond's CUBE-Lite Brings High-Bandwidth Memory to Power-Strapped Edge AI Devices

The memory vendor says its semi-custom CUBE-Lite part delivers up to 8GB/s at roughly 70mW for AI-ISP, wearable and MCU-based designs.. Winbond Electronics announced CUBE-Lite, a semi-custom memory solution aimed at edge AI systems where bandwidth and power are both tight.

InteliToys Editors ? Winbond CUBE-Lite
Winbond CUBE-Lite semi-custom memory (image: Winbond)

Winbond Electronics Corporation announced CUBE-Lite, a semi-custom memory solution built on its customized CUBE platform and aimed at edge AI designs where bandwidth and power are both in short supply. The Taiwan-based semiconductor memory supplier said the part is intended for AI-ISP, wearable, MCU-based and other embedded AI systems, and that it is currently in product validation with mass production targeted for the first half of 2027. The announcement was made at the company's headquarters in Taichung, Taiwan.

According to Winbond, CUBE-Lite delivers up to 8GB/s of memory bandwidth at approximately 70mW, with energy efficiency of roughly 1pJ/bit and an access latency of about 20ns. The company framed those numbers as a direct response to AI processing moving onto endpoint devices, where MCUs, MPUs and NPUs need greater memory bandwidth to support AI workloads while power consumption remains a critical design constraint on battery-powered and thermally limited hardware.

Winbond says the part's simplified interface architecture is its main differentiator, offering an alternative approach to memory integration for edge AI SoCs. The company's engineering data, as cited in its announcement, indicates the interface can reduce SoC controller power consumption by more than 10x compared with an LPDDR4x implementation. Winbond adds that the simplified interface can also reduce the associated PHY die area and SoC integration requirements, giving developers more flexibility to reach the memory bandwidth that AI processing demands while keeping the part practical to integrate into compact endpoint designs.

Ride the Las Vegas Monorail to CES - stops at the Las Vegas Convention Center

Winbond also described the CUBE platform behind CUBE-Lite as a patented high-bandwidth memory interface technology, saying it lets memory modules be optimized for running AI models whose bandwidth demands outstrip what conventional memory modules provide, while reducing power consumption at the same time.

On voltages, Winbond states that CUBE-Lite supports VDD1 at 1.8V, VDD2 at 1.1V and VDDQ from 0.9V down to 0.75V, which the company describes as supporting the low operating voltages used by low-power AI SoCs.

"CUBE-Lite brings the benefits of our CUBE technology to a wider range of Edge AI applications," Winbond said in the announcement. The company positioned the semi-custom approach as a lower entry barrier for customers seeking high memory bandwidth and very low power consumption, while still allowing the solution to align closely with the requirements of applications such as AI-ISP, wearables and MCU-based devices.

Winbond said CUBE-Lite initially targets AI-ISP, wearable and MCU applications, as well as low-power edge vision and embedded AI systems running Micro Linux and AI models. The company added that the semi-custom model lets it work with customers on memory configurations matched to specific edge AI products rather than pushing a single fixed part. Winbond was founded in 1987, is listed on the Taiwan Stock Exchange, and operates 12-inch wafer fabs in Taichung and Kaohsiung.

Source: Winbond.

InteliToys is editorial. We do not sell Winbond CUBE-Lite. Watch the source tape in the branded lightbox below.