A bit technical, but worth reading for anyone tracking $SanDisk Corp.(SNDK)$ and SKHY. The idea is that HBM-HBF combos could be the way forward.
The bottleneck right now is that current GPUs using HBM3e top out at 96GB to 288GB. That forces really large LLMs to either split across hundreds of GPUs or offload data to slower RAM and SSDs.
The tech side draws from storage-inspired memory, applying vertical 3D-stacking techniques similar to 3D NAND flash to high-density DRAM and storage-class memory.
The breakthrough here is stacking memory 128+ layers high, which could scale on-die GPU capacity to multiple terabytes per chip (1TB to 8TB+), at a much lower cost per gigabyte.
The architecture would be tiered. Ultra-fast HBM handles active matrix operations, while the multi-TB 3D DRAM holds entire model parameters with near-nanosecond access.
The impact would be enabling single-node workstation training for 100B+ parameter AI models, much larger context windows, and lower datacenter power and hardware costs.
DRAM.X $Micron Technology(MU)$ $Direxion Daily MSCI South Korea Bull 3x Shares(KORU)$
Disclaimer: Investing carries risk. This is not financial advice. The above content should not be regarded as an offer, recommendation, or solicitation on acquiring or disposing of any financial products, any associated discussions, comments, or posts by author or other users should not be considered as such either. It is solely for general information purpose only, which does not consider your own investment objectives, financial situations or needs. TTM assumes no responsibility or warranty for the accuracy and completeness of the information, investors should do their own research and may seek professional advice before investing.
