Raja Koduri Joins SanDisk’s High-Bandwidth Flash Memory Advisory Board to Achieve 8X-16X Capacity for AI GPUs at 4 TB, Cost-Effective Compared to HBM

Raja Koduri Joins SanDisk’s High-Bandwidth Flash Memory Advisory Board to Achieve 8X-16X Capacity for AI GPUs at 4 TB, Cost-Effective Compared to HBM

Raja Koduri has taken on a pivotal role at SanDisk, aiming to steer the development of High Bandwidth Flash (HBF) memory. This innovation is designed to enhance artificial intelligence (AI) capabilities by addressing the limitations of traditional High Bandwidth Memory (HBM), which often struggles with capacity constraints.

Raja Koduri’s Strategic Appointment to SanDisk’s HBF Memory Advisory Board

Former Intel Chief Architect Raja Koduri, who retired from the tech giant’s Graphics division in 2023, is now dedicated to elevating VRAM capacity in AI GPUs. His recent announcement of joining SanDisk’s Technical Advisory Board highlights his commitment to advancing HBF memory technology, which promises significant improvements in memory capacity for next-gen AI GPUs.

Raja’s collaboration with SanDisk is noteworthy due to his extensive experience in GPU development and computational architectures. This synergy aligns well with SanDisk’s quest to develop HBF technology that overcomes the inherent drawbacks of HBM.

High Bandwidth Flash stack diagram enhancing HBM memory with NAND flash for AI workloads.
The HBF stack can deliver multiple times greater memory capacity while maintaining the same bandwidth as HBM.

When we began HBM development our focus was improving bandwidth/watt and bandwidth/mm² (both important constraints for mobile), while maintaining competitive capacity with the incumbent solutions. With HBF the focus is to increase memory capacity (per-dollar, per-watt, and per-mm²) significantly while delivering competitive bandwidth.

Raja Koduri

Despite the rapid advancements in HBM, which has successfully offered large memory capacities for AI-centric superchips, HBF has the potential to exponentially increase memory capabilities by utilizing Through-Silicon Vias technology. A single HBF stack can enable terabyte-level memory capacity, and integrating eight such stacks into a system can propel AI GPUs to achieve up to 4 TB of VRAM while retaining high bandwidth characteristics that HBM provides. This development is crucial to meeting the escalating demands of AI applications.

It’s essential to understand that SanDisk’s HBF technology will not directly compete with DRAM in latency-sensitive tasks. Instead, it is tailored to meet the memory requirements of AI operations, such as inference and large-scale model training, which prioritize capacity and bandwidth over latency. Raja’s strategic role will be instrumental in advancing this high-capacity memory solution.

Comparison between HBM and HBF GPU memory capacities.Visual representation of GPU memory capabilities: 192GB HBM vs 4, 096GB HBF.Running Frontier LLM with HBF: 1.8T parameters, 16-bit weights, and 3, 600GB memory for GPU.

As SanDisk aims to establish HBF as an open-standard framework, this initiative is expected to foster widespread adoption across the industry. Raja’s remarkable network and experience in building ecosystems will be invaluable in enhancing collaborations with GPU manufacturers.

HBF is set to revolutionize edge AI by equipping devices with memory capacity and bandwidth capabilities that will support sophisticated models running locally in real time. This advancement will unlock a new era of intelligent edge applications, fundamentally changing how and where AI inference is performed.

– Raja Koduri

For more details, visit the official announcement from SanDisk.

Source & Images

Leave a Reply

Your email address will not be published. Required fields are marked *