NVIDIA Uninterested in HBF Memory Technology Despite New 4TB Stacks Surpassing HBM, Google Begins Sampling This Year

High-Bandwidth Flash (HBF) memory is emerging as a notable contender in the memory technology landscape, boasting greater capacity than High Bandwidth Memory (HBM).However, NVIDIA has opted not to adopt this innovative solution; instead, Google is set to become a primary customer for HBF.

NVIDIA Remains Committed to HBM: HBF Sampling Planned for Later This Year

Recent advancements in NAND DRAM come at a time when artificial intelligence (AI) applications are rapidly gaining momentum. Although NAND primarily serves storage needs—primarily in solid-state drives (SSDs)—the anticipated advancements in HBF technology could significantly transform memory solutions. Designed to bridge HBM and NAND Flash, HBF represents the next generation of NAND DRAM.

The architecture of HBF will incorporate numerous Through Silicon Vias (TSVs), allowing for the integration of multiple NAND packages into a single stack. Currently, HBM supports capacities ranging between 32 to 64 GB per stack, while HBF is poised to offer expansive capacities of up to 4 TB.

In terms of performance, while HBM retains its status as the faster option, architectural optimizations in HBF are expected to yield sufficient throughput for essential AI tasks. This new standard is particularly well-suited for inferencing workloads, which have surged in importance alongside the rise of Agentic AI. The increased capacity offered by HBF may also help alleviate some of the constraints imposed by Key-Value (KV) Caches in primary compute chips.

NAND related update: NVIDIA still does not appear interested in HBF. Its view is that HBF’s high bandwidth can be sufficiently addressed with eSSDs. There were recent reports that SanDisk had issued POs related to HBF, and the customer SanDisk is believed to be in discussions…

— Jukan (@jukan05) April 28, 2026

Despite the potential advantages of HBF, NVIDIA has publicly stated that it has no immediate plans to implement this new DRAM technology, believing that existing enhanced SSDs (eSSDs) can adequately meet current capacity and speed requirements. The company is collaborating with Kioxia to develop PCIe Gen7 SSDs that could deliver speeds up to 100 times faster than conventional models.

A SanDisk presentation slide titled 'High Bandwidth Flash (HBF™)' details augmenting HBM memory with NAND flash for AI workloads, featuring a diagram of the HBF stack with components labeled as HBF Core Die, Logic Die, PHY, and Interposer. — Image Source: SanDisk

In contrast, Google appears ready to capitalize on HBF technology as part of its ambitious AI expansion strategy. The tech giant’s Tensor Processing Unit (TPU) ecosystem is rapidly evolving, with a pipeline of next-generation TPU solutions aimed at enhancing computational capabilities. While the prominence of HBF in broader applications remains uncertain, its potential to replace standard DDR memory marks an exciting development in the field.

As computing demands increase, servers are increasingly adopting Low Power Double Data Rate (LPDDR) memory, driven by the constraints presented by CPUs in AI applications. This trend has highlighted the growing need for LPDDR5 and LPDDR5X memory, particularly for System-on-Chip and Multi-Chip Module (SOCAMM2) configurations. HBF’s innovative multi-layer stacking approach enables chip manufacturers and AI ecosystem stakeholders to reduce printed circuit board (PCB) size while increasing capacity, maintaining low power consumption, and achieving high bandwidth throughput.

Source & Images