NVMe Hard Drives: Seagate's Answer to Growing AI Storage Demands

[Nomad76]

Seagate is advancing in NVMe technology for large-capacity hard drives as the growing storage needs in AI systems and apps continue to rise. Old and current storage setups struggle to handle the huge datasets needed for machine learning, which now reach petabyte and even exabyte sizes. Today's storage options have many drawbacks for AI tasks. SSD-based systems offer the needed speed however it is not cost-effective for storing large AI training data. SAS/SATA hard drives are cheaper however, the interfaces rely on proprietary silicon, host bus adapters (HBAs), and controller architectures that struggle with the high-throughput, low-latency needs of AI's high data flow.

Seagate's "new idea" is to use NVMe technology in large-capacity hard drives. This creates a solution that keeps hard drives cost-effective and dense while boosting performance for AI apps. NVMe hard drives will not need HBAs, protocol bridges, and additional SAS infrastructure, allowing seamless scalability by integrating high-density hard drive storage with high-speed SSD caching in a unified NVMe architecture.



Moreover, NVMe hard drives offer several benefits. They're easy to set up without special controllers, and they can be mixed with SSDs using a single NVMe driver and OS stack. Another advantage is the support for direct data transfer from GPU to storage using data processing units (DPUs), which avoids CPU bottlenecks. Also, NVMe over Fabrics (NVMe-oF) allows NVMe hard drives to integrate into distributed AI storage architectures.

To test this technology, Seagate ran a proof of concept (POC) integrating NVMe hard drives, NVMe SSDs, NVIDIA BlueField DPUs, and AIStore software (8x NVMe hard drives and 4x NVMe SSDs, all directly connected to NVIDIA's BlueField 3 DPU). The results showed a lower storage-related latency, a simplified infrastructure setup, better caching, and storage scalability. Seagate demonstrated that NVMe hard drives don't just work better. They're also much greener. They use ten times less embodied carbon per terabyte and four times less power than SSDs. Plus, they cost way less per terabyte.

Compared to SSDs, NVMe hard drives would offer:

• 10× more efficient embodied carbon per terabyte, significantly reducing environmental impact.
• 4× more efficient operating power consumption per terabyte, lowering AI data center energy costs.
• Significantly lower cost per terabyte, reducing AI storage TCO at scale.

Seagate has big plans as they want to make their Mozaic platform support more than the current 36 TB drives. They plan to improve NVMe over Fabrics (NVMe-oF) support and are currently working with their customers and partners to explore how NVMe hard drives can fit into next-generation AI storage solutions. Seagate admits that HDDs can't compete with solid-state drives when it comes to pure speed. However, their upgraded NVMe HDDs can provide a more efficient storage option for next-gen data centers.

View at TechPowerUp Main Site | Source

 

[R-T-B]

I called it when linux kernel 6.13 added a commit for this, knew it was incoming.

Unless they are benching against optane or something, I'm calling BS on that "greener than SSDs" claim.

 

[Assimilator]

Wait, is this just spinning rust connected over PCIe (NVMe) instead of SATA/SAS? If so, why? Is this Seagate just trying to keep HDDs relevant with marketing BS?

 

[R-T-B]

Wait, is this just spinning rust connected over PCIe (NVMe) instead of SATA/SAS? If so, why? Is this Seagate just trying to keep HDDs relevant with marketing BS?

Faster access to big cache. Think they are trying to create an enterprise version of "hybrid drives."

 

[cal5582]

honestly hybrid drives were a good idea. slap a 1tb nvme on a huge drive as cache and go.

 

[Assimilator]

Faster access to big cache. Think they are trying to create an enterprise version of "hybrid drives."

That's what I assumed, but the block diagram they presented doesn't appear to be congruent with that; I would expect the "mass-capacity hard drive storage" and "SSD cache" to be represented as a single block if it was a hybrid HDD situation, but they're not...

 

[Nomad76]

That's what I assumed, but the block diagram they presented doesn't appear to be congruent with that; I would expect the "mass-capacity hard drive storage" and "SSD cache" to be represented as a single block if it was a hybrid HDD situation, but they're not...

To my understanding, the SSDs are for active AI data, HDDs will store the large training data sets. Cost reduction came also by interposing/interconnecting the DPU as a single unified node. Not sure it will fit all scenarios but I think that in some fields i.e. research, it can be a good $$$ option.

"By using NVMe hard drives alongside SSDs, organizations will be able to optimize cost while maintaining performance, reserving SSDs for active datasets and using hard drives for long-term AI training data retention.", Seagate

 

[Scrizz]

Wait, is this just spinning rust connected over PCIe (NVMe) instead of SATA/SAS? If so, why? Is this Seagate just trying to keep HDDs relevant with marketing BS?

That's exactly what it is. Someone wants to access everything over the same bus and manage things the same way.
I wonder if NVMe Optical drives exist.

 

[framebuffer]

maybe now people will stop using NVMe as a synonym for SSD

 

[R-T-B]

I wonder if NVMe Optical drives exist.

No but you can pump nvme over different fabrics including but not limited too, optical. Close enough?

 

[ZeDestructor]

That's exactly what it is. Someone wants to access everything over the same bus and manage things the same way.
I wonder if NVMe Optical drives exist.

Not yet. Personally, I think they'll go to USB-native controllers rather than NVMe.

 

[Scrizz]

maybe now people will stop using NVMe as a synonym for SSD

I hope so, but now they'll say PCIe drives