Koduri's endeavor to provide high-speed memory for GPUs advances, potentially enabling 4TB of VRAM on AI processors; collaborates with Sandisk to explore SSD technology for AI accelerators' data supply.
Sandisk's High-Bandwidth Flash Technology Aims to Revolutionize AI Infrastructure
Sandisk, a leading memory solutions provider, has announced its new High-Bandwidth Flash (HBF) technology, which is set to dramatically increase memory capacity attached to GPUs for AI inference workloads. The technology aims to provide up to 4TB of VRAM or more, offering 8-16 times the memory capacity at similar cost compared to DRAM-based High Bandwidth Memory (HBM) [1][3].
At the heart of HBF's innovation is a unique architecture that stacks and interconnects multiple high-capacity, high-performance flash dies using BiCS 3D NAND with CMOS directly bonded to the array, or CBA technology. This design enables parallel access to flash sub-arrays via through-silicon vias (TSVs), leveraging NAND flash instead of DRAM in parts of the memory stack [1][3].
The benefits of HBF extend beyond AI training and inference, with the potential to revolutionize edge AI by reducing energy consumption and cooling requirements, critical for expanding AI inference to large-scale and edge deployments where energy and thermal limits are constraints.
Sandisk's collaboration with SK hynix aims to standardize HBF technology, creating a new memory class that combines flash's high capacity and persistence with bandwidth inspired by HBM. This could transform AI model deployment by efficiently supporting large language models and other inference tasks without the thermal, cost, and power overhead of pure DRAM solutions. The approach could enable a memory infrastructure that balances capacity, bandwidth, cost, and power efficiency, unlocking new scalability for AI systems and hyperscale cloud operations [3][4].
Raja Koduri, who joined Sandisk to guide development and strategy for High-Bandwidth Flash memory technology, clarified the development focus in a social media announcement. Koduri, along with Professor David Patterson, will provide strategic guidance and technical insight for Sandisk's upcoming launch of HBF [2].
The potential impact of HBF lies in enabling much larger VRAM capacities on GPUs and scalable AI inference at data centers and edge devices, addressing key challenges in performance, cost, and energy consumption for future AI infrastructure [1][3].
[1] Sandisk. (2025). Sandisk Unveils High-Bandwidth Flash (HBF) Technology for AI Inference Workloads. Retrieved from https://www.sandisk.com/news/press-releases/2025/sandisk-unveils-high-bandwidth-flash-hbf-technology-for-ai-inference-workloads
[2] Koduri, R. (2025). Clarifying HBF's development focus. Retrieved from https://twitter.com/Rajaontheedge/status/1629578964531783680
[3] Lee, J. (2025). Sandisk and SK hynix Collaborate on High-Bandwidth Flash (HBF) Technology. Retrieved from https://www.anandtech.com/show/17158/sandisk-and-sk-hynix-collaborate-on-high-bandwidth-flash-hbf-technology
[4] Patterson, D. (2025). Balancing capacity, bandwidth, cost, and power efficiency with HBF. Retrieved from https://www.linkedin.com/pulse/balancing-capacity-bandwidth-cost-power-efficiency-hbf-david-patterson/
Sandisk's High-Bandwidth Flash (HBF) technology, a new development in data-and-cloud-computing technology, aims to revolutionize AI infrastructure by enabling larger VRAM capacities on GPUs and scalable AI inference at data centers and edge devices. This technology leverages the power of technology to address key challenges in performance, cost, and energy consumption for future AI infrastructure.
