The big picture: With memory prices skyrocketing, tech companies are exploring new ways to reduce the cost of AI development. Earlier this year, Google detailed its TurboQuant compression technique, which reportedly reduces LLM memory usage by up to 6×. Now, researchers in Belgium have developed a NAND-DRAM hybrid architecture that they claim could significantly lower AI inference costs in the future.
Leuven, Belgium-based nanotechnology and semiconductor research center Imec has unveiled what it describes as the world's first 3D implementation of a charge-coupled device (CCD) designed for AI memory applications. The new technology combines the speed of DRAM with the storage density of NAND flash, potentially reducing the "memory wall" bottleneck, in which limited memory bandwidth forces AI accelerators to wait for data instead of processing tokens continuously.
The device is built by stacking memory chips vertically rather than placing them side by side, enabling ultra-fast charge transfer speeds reportedly exceeding 4GHz under laboratory conditions. To reduce leakage and support denser 3D integration, the researchers used indium gallium zinc oxide (IGZO), a compound that offers significantly better electron mobility, energy efficiency, and optical transparency than traditional silicon.
As noted by TechRadar, CCD technology was once widely used in digital cameras, broadcast video equipment, scientific imaging devices, and astronomy sensors. However, it has largely been replaced by CMOS sensors, which are faster, more power-efficient, and less expensive to manufacture. CMOS sensors also support on-chip integration of key functions such as analog-to-digital conversion and image processing, enabling slimmer and more efficient camera designs.
It is worth noting that 3D CCD technology remains in the proof-of-concept stage, meaning substantially more research is needed to determine whether it can be scaled reliably, efficiently, and economically for real-world applications. The researchers believe the technology is unlikely to appear in data center servers anytime soon. However, if issues related to thermal behavior and layer scaling can be addressed, it could eventually absorb some of the growing demand for DRAM and HBM in AI data centers and other electronics markets.
3D CCD is the latest attempt by researchers to develop alternative solutions for soaring memory prices. In March, Google announced three AI compression algorithms designed to significantly reduce the memory footprint of LLMs without degrading output quality. According to the company, TurboQuant, PolarQuant, and Quantized Johnson-Lindenstrauss can shrink model sizes with "zero accuracy loss," improving vector search efficiency and reducing key-value cache bottlenecks.