Overview
- The companies released the HBF technical spec through the Open Compute Project on Tuesday, defining up to 512 GB per package and three bandwidth grades from roughly 0.4 TB/s to 3.0 TB/s.
- The spec adopts the Universal Chiplet Interconnect Express (UCIe) interface so HBF devices can sit close to CPUs and GPUs and be integrated into chiplet-based systems.
- SK hynix and Sandisk position HBF as a middle memory tier between HBM and SSDs to give AI inference systems more capacity near the accelerator and reduce data-transfer bottlenecks.
- Only Google and Tenstorrent are named consortium members so far, and major players such as Nvidia, AMD, Samsung and Micron have not publicly joined, leaving real-world uptake unproven.
- Commercial samples and early products are expected over the next 6–18 months while competing approaches from Samsung and others and a tight memory supply market will shape who adopts HBF and how quickly it reaches data centers.