Ahead-looking: Onerous disk drive expertise dates again to the Fifties and nonetheless depends on decades-old I/O requirements. Seagate plans to usher arduous drives right into a PCIe-based future – a shift pushed by the rising demand for sooner storage in synthetic intelligence coaching and inferencing, notably in large, energy-hungry knowledge facilities.
Seagate lately launched a novel method to HDD growth and evolution, bringing magnetic storage and NVMe collectively in a brand new storage expertise. The producer desires to simplify and enhance knowledge storage for AI firms by offering a considerably cheaper answer than solid-state drives.
Onerous disk drives won’t ever attain the identical efficiency stage as SSDs, however Seagate says it may enhance them for extra environment friendly use in next-gen knowledge facilities. The corporate is pioneering a “transformational” answer by including NVMe compatibility to high-capacity arduous drives.
Seagate defined that by adopting the PCIe-based protocol, NVMe arduous drives might get rid of the necessity for proprietary silicon tied to conventional SAS/SATA interfaces, host bus adapters (HBAs), and controller architectures unsuitable for AI workloads. The brand new protocol guarantees excessive throughput and low latency, although it is arduous to think about a future the place arduous disk drives transfer knowledge at gigabytes per second.
These NVMe arduous drives may have no HBAs or customized controllers, however they’ll retain the identical SAS/SATA connector for improved compatibility. The brand new tech will simplify the deployment of AI storage options, whereas a single NVMe software program driver stack ought to make HDDs and SSDs work collectively extra effectively.
Seagate can be touting the power to immediately join NVMe drives to GPUs, bypassing the eventual bottlenecks created by conventional knowledge routes via the CPU. The brand new structure is extra environment friendly, so AI fashions can course of large knowledge units with decreased I/O delays. Moreover, a novel NVMe over Materials (NVMe-oF) expertise ought to carry HDDs into totally distributed, scalable storage architectures.
Seagate examined the potential influence of NVMe arduous drives via a proof-of-concept integration between customized HDDs, NVMe SSDs, Nvidia BlueField knowledge processing models (DPU), and the AIStore software program. Whereas the corporate hasn’t supplied particular metrics concerning efficiency, it claims that the brand new platform supplied some key benefits for AI workflows. Direct GPU-to-storage communication by way of DPUs decreased storage-related latencies whereas eliminating legacy SAS/SATA overhead within the less complicated system structure.
Moreover, the AIStore software program helped optimize knowledge caching for higher AI mannequin coaching efficiency, whereas NVMe-oF integration proved its usefulness in multi-rack storage clusters. Seagate stated that NVMe HDDs would provide 10 instances extra environment friendly embodied carbon per terabyte, 4 instances extra environment friendly energy consumption, and “considerably decrease” price per terabyte than SSDs. The corporate is now working to scale its HAMR-based Mozaic 3+ platform and develop higher-capacity arduous disk drives whereas exploring NVMe storage options with its companions.
