Seagate has unveiled a groundbreaking storage solution that brings NVMe protocol support to traditional hard disk drives (HDDs), potentially transforming how AI data centers handle massive datasets. The company recently demonstrated a proof-of-concept system that pairs HDDs with GPU direct access, marking a major shift in storage technology.
The new system combines eight NVMe HDDs and four NVMe SSDs for caching, powered by Nvidia's BlueField 3 data processing unit (DPU) and AIStore software. This setup enables direct GPU-to-storage access, bypassing CPUs and reducing system bottlenecks.
Traditional hard drives use SATA and SAS interfaces, which were developed in the 1980s and carry outdated protocol layers. These legacy systems require additional hardware like host bus adapters, creating complexity and latency issues that hinder AI workload performance.
By adopting the NVMe protocol over PCIe, Seagate's new drives can achieve dramatically higher bandwidth - up to 128 GB/s at the system level - while simplifying the overall architecture. The technology supports 64K queues with 64K commands per queue, substantially improving parallel processing capabilities needed for AI applications.
The demonstration showed that NVMe HDDs could maintain high performance for AI environments without requiring expensive all-flash storage solutions. The system can scale to exabyte levels using NVMe-over-Fabric technology, allowing seamless expansion across large data center installations.
The cost impact appears minimal, as the HDDs retain their traditional physical connectors and form factors. The main changes involve adding NVMe protocol support and PCIe interface to the controller, along with new firmware development.
While Seagate leads in this innovation, other hard drive manufacturers are also developing NVMe-capable drives as part of the Open Compute Project. Mass adoption is expected once multiple vendors can supply these drives at scale, particularly targeting cloud service providers focused on AI workloads.
This advancement comes at a critical time as enterprises increasingly manage massive datasets for AI training and inference, requiring more efficient and cost-effective storage solutions that can handle the growing demands of artificial intelligence applications.